Unlike traditional AI backdoors, which manipulate a model's learned parameters, architectural backdoors hide malicious instructions directly in the code that defines how the model operates.
However, everyday image-processing operations, such as cropping, adjusting contrast, or adding noise, can change a trigger pattern and prevent an attack from working, according to new research from Georgia Tech.
The findings challenge assumptions about how practical these attacks are. Although previous research showed that architectural backdoors could work under controlled conditions, the Georgia Tech team found that existing approaches struggle to remain effective and difficult to detect when AI systems process images under realistic conditions.
“The biggest misconception is that demonstrating an architectural backdoor under controlled conditions means it will actually work in the real world,” said David Oygenblik, Ph.D. student and lead author of the study.
The study details the development of FAULT, a framework for evaluating architectural backdoors under realistic conditions.
The findings reveal a tradeoff. Attackers can make a backdoor more sensitive, so it recognizes its trigger after an image has been changed. But that also increases the chance that ordinary images will accidentally activate the backdoor, potentially hurting the AI system’s performance.
“For the architectural backdoors we study, this tradeoff is fundamental,” Oygenblik said.
The team found that this problem persisted across five methods of implementing architectural backdoors. Even when attackers knew how the system would process images, they struggled to make their attacks both reliable and difficult to detect.
The research did not test attacks against real vehicles or other production systems, nor did it demonstrate outcomes such as crashes or ransomware. Instead, it examined whether existing architectural backdoors could reliably activate under different image-processing conditions.
However, the findings highlight a potential risk in the AI software supply chain. If an attacker inserted malicious code into an AI model used in a real system, the backdoor could potentially cause incorrect predictions when its trigger was detected. The consequences would depend on how the system used the model’s output.
Because architectural backdoors hide in code, examining a model’s learned parameters or retraining it may not be enough to uncover or remove the threat.
Oygenblik said developers should test AI systems under realistic conditions, watch for unexpected drops in accuracy, and examine model code for suspicious trigger-detection logic.
Finding FAULTs in Architectural Backdoors: Why Trigger Detection Fails Under Real-World Conditions, was accepted to the ACM Conference on Computer and Communications Security. The authors will present the paper in The Hague in November.
Oygenblik conducted the research with undergraduate students Teja Akella and Tanmay Gupta, Ph.D. student Yizhi Huang, and Associate Professor Brendan Saltaformaggio of Georgia Tech’s Cyber Forensics Innovation Laboratory. The lab is a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering.
]]>• Cropping, contrast changes, and similar transformations can disrupt visual triggers across five backdoor methods.
• Increasing trigger sensitivity improves reliability but also raises accidental activations and may reduce system performance.
• The study warns that parameter inspection or retraining may miss code-based threats, urging realistic testing and code reviews for suspicious trigger logic.
Communications Officer for the School of Cybersecurity and Privacy
]]>Georgia Tech researchers have developed a tool that can recover deleted conversations from AI assistants running on Android phones. The tool, called ORISA, can also rebuild the recovered information into a working AI session.
The findings show that deleting an AI conversation on a phone does not necessarily mean it removes every trace.
The research could give digital investigators a new way to recover information from phones. It also raises privacy concerns for people who use AI assistants that run locally on their devices.
Local AI assistants run directly on a phone rather than sending conversations to a cloud service. While the local assistant is running, information from conversations can remain in the phone’s working memory even after it is deleted.
ORISA searches this memory for pieces of deleted conversations.
In one experiment, researchers created a fictional AI assistant called “Project Note Keeper” using the ChatterUI Android app. They gave it 100 fictional facts about a project, then deleted the conversation.
ORISA recovered information from the deleted conversation and used it to create a new working session. When researchers asked the restored assistant to list the project notes it had been given, it returned 26 facts. The original session could access only 14 facts because the rest had fallen outside its active memory.
The experiment showed that researchers could recover information the AI had already stopped actively using.
“It can help establish what information someone shared with an AI assistant and what assistance they received,” said Ph.D. student Ken Xu, lead author of the study.
The researchers tested ORISA with 10 Android apps and three AI models. Across 60 test configurations, the tool recovered conversation information that remained in the phone’s memory.
The researchers say the information becomes harder to recover as the phone’s memory is overwritten by other activity.
The research was presented at the 2026 IEEE Symposium on Security and Privacy. The paper, titled Recovering and Rehosting Mobile Local LLM Conversations and Contexts via Memory Forensics, was authored by Ph.D. students Xu, David Oygenblik, Runze Zhang, and Mingxuan Yao; Research Scientist Muhammad Ibrahim; and Associate Professor Brendan Saltaformaggio. The work was done in Georgia Tech's Cyber Forensics Innovation Laboratory, a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering.
]]>However, a new paper by a Georgia Tech-led research team suggests that the requirements can disproportionately burden family farms and other small food producers due to a lack of technological infrastructure, resources, or compliance expertise.
The rule is mandated by Section 204(d) of the Food Safety Modernization Act, which Congress passed and President Barack Obama signed into law in 2011. Intended to help the FDA quickly identify the source of contamination, the rule requires producers or distributors with more than $25,000 in annual produce sales to comply or face significant fines.
Earlier this year, the FDA extended its compliance deadline to July 2028. One reason for the extension may be the volume of feedback from stakeholders affected by the rule. The FDA received more than 1,200 comments during an open forum, with feedback from stakeholders ranging from major distributors to small farmers.
“Many small farmers commented on how they may not have the devices needed to input this data,” said Soonho Kwon, a Ph.D. student in Georgia Tech’s School of Interactive Computing (SIC) and lead author of the paper. “Does this mean they have to purchase it themselves? Do they need to become educated on how to use computing systems?”
Kwon is a member of the Equity and Sustainability in Computing Lab, which is led by SIC professor and chair Shaowen Bardzell.
Many small farms also said they can’t afford to hire someone to track the data.
One small farmer said the amount of data tracking required by the rule is overwhelming. Some of the required data include:
“Please do not force us to do technology stuff,” the commenter said. “It’s hard enough to keep up with the local demand as it is. We run a pretty low digital technology effort. It’s basically harvest fresh organically grown veggies and sell them to local people.”
Kwon said the intention of the Food Traceability Rule is admirable.
“Preventing foodborne illness is critical, and data- and technology-driven governance can be a powerful tool toward that goal,” he said. “But it’s equally important to think critically about who is supposed to be the default user of these systems, and who gets missed in the process. It’s undermining the people who are unable to comply.”
Other factors didn’t appear to be considered by the FDA, including the lack of technological infrastructure in rural America, where many foods are grown.
“Many parts of rural America have limited or unreliable internet access,” Kwon said. “If the FDA wants them to send this data, they must comply within 24 hours. Does that mean they have to stop their daily routine and drive to a city that has internet to send it? The rule is operating on the assumption that everyone is dealing with the same physical environment.”
Culture and religion also seem overlooked. Some commenters raised concerns that Amish farmers who avoid certain technologies for religious reasons would be pressured into digital record keeping.
Many major distributors and organizations expressed support for the Food Traceability Rule in the FDA’s forum. Catherine Wieczorek, a Ph.D. student in SIC and co-author on the paper, said these organizations don’t have the same tension points as smaller producers.
“Larger food producers have the operational capacity to do the type of work the rule requires,” Wieczorek said. “We find there’s a power imbalance when it comes to smaller food producers like a family farm. Change at that scale is difficult for smaller producers.”
Bardzell, who advises Kwon and Wieczorek, said the situation provides computing researchers an opportunity to create interventions that will benefit those with fewer resources.
“In a data age when we’re adding AI into our work processes and labor practices, we need to think about how this need creates unequal burdens on different stakeholders in our food system,” Bardzell said. “It’s a significant challenge for computing going forward.”
Kwon will present the paper on behalf of his collaborators next week at the ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW) in Salt Lake City, Utah. Their research is supported by the National Science Foundation.
]]>In a study accepted for the 2026 ACM Conference on Computer and Communications Security, Georgia Tech researchers interviewed 20 military sailors about their cybersecurity experiences, training, and responses to hypothetical scenarios. Their findings point to a gap between how mariners are trained to handle emergencies and how cyberattacks can compromise the systems they depend on.
The study found several gaps in how crews prepare for cyber threats:
These gaps matter because cyberattacks can compromise a system without causing it to stop working. For example, an attacker could manipulate navigation data, leaving a crew to make decisions based on false information.
“I’d assume it was anything but cyber at first. It would probably take me quite some time to get to that conclusion,” one participant told the researchers.
The team examined three hypothetical scenarios: GPS spoofing, loss of propulsion control in a high-traffic area, and failure of electronic navigation systems in a congested channel at night.
In each scenario, mariners would first focus on keeping the vessel safe, using established procedures to stabilize the ship and troubleshoot its systems. While this response is essential during an emergency, it could delay recognition of an attack if crews do not consider that the system or its data may have been compromised.
For example, a ship navigating a narrow channel depends on accurate location information, but what happens when its navigation systems appear to work normally while a hacker manipulates that data?
An attacker could use GPS spoofing to falsify a vessel’s reported position. If the false position is close enough to the ship’s actual location, a mariner might not notice anything wrong. The vessel could gradually head toward shallow water or another obstacle without triggering an obvious error or alarm.
“By the time the mariner realizes something is wrong, the ship’s size and limited maneuverability could make it too late to avoid a grounding or collision,” said Anna Raymaker, a Georgia Tech Ph.D. student who studies maritime security.
The scenario illustrates a central challenge in maritime cybersecurity: An equipment failure may be easy to recognize, but a cyberattack that provides false information can be much harder to detect.
Batten the Hatches: Cybersecurity with Military Mariners was written by master’s student Ryan Von Brock, Raymaker, Research Scientist Animesh Chhotaray, Associate Professors Frank Li and Saman Zonouz, and Provost & Executive Vice President for Academic Affairs Raheem Beyah.
The research was conducted in Georgia Tech’s Cyber-Physical Security Lab, a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering.
Photos by Terence Rushin/College of Computing
]]>New research from Georgia Tech suggests that a simple change to a chatbot’s instructions could reduce that risk. The work was accepted to the ACM AI Letters publication.
Researchers Lorenzo de la Loza and Vijay Madisetti tested eight AI models in 1,280 experiments involving simulated conversations about delusional beliefs. They found that instructions designed to discourage excessive agreement reduced the models’ confirmation of those beliefs by 73.8%. Adding a self-reflection step increased the reduction to 77%.
“We think that this paper will help AI providers and solution creators better understand how best to protect and safeguard their users from harming themselves,” Madisetti said. “This may make AI safer for all of us.”
AI chatbots are designed to be helpful and supportive. But that behavior can become risky when a chatbot agrees with a user’s beliefs without questioning whether they reflect reality. Researchers call this behavior sycophancy.
Courtney Crooks, a principal research scientist at the Georgia Tech Research Institute and adjunct faculty member in Georgia Tech’s School of Cybersecurity and Privacy, said chatbot conversations can have both positive and negative effects.
People may feel more comfortable discussing difficult thoughts with a chatbot than with another person. But if the chatbot always agrees, users may miss opportunities to question their beliefs or consider other points of view.
The same behavior that makes a chatbot feel supportive, Crooks explained, can also encourage harmful patterns of thinking.
The researchers tested instructions that told AI models not to reinforce users’ beliefs simply to be agreeable. The team evaluated the models using 16 clinically derived scenarios involving delusional beliefs.
The team also tested other safety tools, including systems designed to flag potentially harmful content. The results suggest that changing a chatbot’s instructions can be an effective way to reduce excessive agreement in the scenarios tested.
Crooks called the findings a promising direction for future research combining psychology and AI safety.
The study did not test people in clinical settings or prove that the approach prevents real-world harm. However, it offers developers a potential way to make chatbots more careful when responding to vulnerable users without building an entirely new AI model.
]]>Communications Officer for the School of Cybersecurity and Privacy
]]>After a grueling, world-class competition, the team won the $4 million first-place prize in August 2025.
The team has since disbanded. Some members have moved into the corporate world, while others continue in academia. Despite these varied paths, Team Atlanta’s search for breakthrough AI cybersecurity solutions continues.
Team Atlanta lead Taesoo Kim, principal security researcher Dongkwan Kim, and three other former team members joined Microsoft Security after the competition, where they continue to work on AI and cybersecurity in the Frontier Offensive Research & Generative Exploitation (FORGE) Lab.
The move has given the researchers an opportunity to apply concepts from AIxCC to real-world software. It has also exposed them to challenges that are difficult to replicate in a competition or research lab.
Dongkwan Kim said that team members continue to share what they learned through research papers, reports, software, and benchmarks. At Microsoft, these lessons are tested against real customer code.
This work led to MDASH, Microsoft's multi-agent, multi-model system for vulnerability discovery and remediation. The system uses more than 100 AI agents to search for vulnerabilities, validate potential findings, and help develop fixes.
Rather than asking a single AI model if a piece of code contains a vulnerability, MDASH can examine a problem using multiple agents. The system can also try to create an exploit to determine whether a potential vulnerability is real and reachable.
For Dongkwan Kim and the other former Team Atlanta members at Microsoft, moving from AIxCC to MDASH showed how research can look different when applied to real software.
“AIxCC was not an endpoint for Team Atlanta,” he said. “We continue sharing what we learned through technical reports, papers, open artifacts, and benchmarks for vulnerability discovery and patching.”
Working with customer code introduced new challenges for the Microsoft team. The researchers had to consider security boundaries, existing development processes, how findings would be reviewed, and how vulnerabilities would be prioritized for action.
These challenges created a new feedback loop between research and product development. The team took ideas from AIxCC, test them on real software, learn where they worked and where they fell short, and use those lessons to improve the system.
The experience also showed that AI does not replace the need for cybersecurity experts. Human researchers are still needed to provide context, decide how to break down difficult problems, validate findings, understand attack paths, and determine the potential impact of a vulnerability.
Taesoo Kim, a professor in the School of Cybersecurity and Privacy and vice president of security research at Microsoft, said the competition continues to influence the Microsoft team's work.
“Every single activity that we’ve done at Microsoft was influenced by AIxCC,” he said in a 2026 interview with Microsoft.
The work has continued to grow. In 2026, Microsoft introduced MAI-Cyber-1-Flash, an AI model built specifically for cybersecurity and designed to work with MDASH.
The model is designed to handle many of MDASH's tasks while allowing larger AI models to focus on harder problems. Microsoft reported that replacing 80% of the models used by MDASH with MAI-Cyber-1-Flash increased its performance on the CyberGym benchmark from 88.4% to 95.95%.
The benchmark includes more than 1,500 examples from 188 software projects. Microsoft also reported that the new setup can achieve similar performance at half the cost of the previous MDASH configuration.
“AIxCC has sparked cybersecurity innovation across the globe, from frontier AI companies to small academic labs,” said Taesoo Kim. “It is an incredibly exciting time to be working in AI-driven security engineering and research, as we see AI fundamentally reshape how we find and fix vulnerabilities at scale.”
The AIxCC experience also continues to shape cybersecurity research at Georgia Tech.
Jiho Kim, a Georgia Tech Ph.D. student who was part of Team Atlanta's research efforts, said one of the biggest lessons he’s brought back to campus was learning how AI and traditional security tools can work together.
One of these tools is fuzzing. Fuzzing tests software with large numbers of inputs to detect unexpected behavior or security problems. Large language models (LLMs), meanwhile, can reason about code and help guide the search.
During the AIxCC competition, the Team Atlanta combined these ideas in CtxFuzz, a system that uses an LLM to help guide fuzzing toward parts of software that may contain vulnerabilities.
The research was submitted to the ACM/IEEE International Conference on Automated Software Engineering, where it received the ACM Distinguished Paper Award.
“Since AIxCC, we’ve continued building on some of the core findings from our two-year journey with Team Atlanta,” Jiho Kim said.
“One of the biggest lessons for us was understanding how LLMs and traditional fuzzing can work together in a mutually beneficial way, using each to overcome the limitations of the other and ultimately find more and deeper vulnerabilities.”
The researchers also wanted to make the lessons from AIxCC useful to other cybersecurity researchers.
They created an Open-Source Cyber Reasoning System Framework (OSS-CRS) and CRSBench, which provides researchers and security professionals with ways to use, test, and build on technology developed through AIxCC.
“We’ve also been focused on making the results of AIxCC useful to the broader security community,” Kim said. “We created OSS-CRS and CRSBench so that researchers and practitioners can more easily use, reproduce, and build upon the capabilities and lessons from our AIxCC work.”
The team's research has also moved beyond identifying vulnerabilities. Researchers are studying how well AI models can understand and reason about computer code.
Jiho Kim and his lab mates in the Systems Software & Security Lab developed COIN, a benchmark designed to test how well LLMs can reason about code. The work has been accepted to the 2026 Conference on Empirical Methods in Natural Language Processing.
The Georgia Tech researchers are also developing additional benchmarks to study the various types of reasoning required for cybersecurity. They are using what they learn to create new datasets and methods for training AI models focused on cybersecurity.
“We’re using what we learn from these benchmarks to train cybersecurity-focused models through new datasets and training methodologies,” Jiho Kim said.
Photos by John Popham/College of Computing
]]>Communications Officer at the School of Cybersecurity and Privacy
]]>Those conversations led Raymaker to focus her research on maritime security, including vulnerabilities in the Automatic Identification System (AIS), which ships use to broadcast their location, speed, and course.
This summer, Raymaker brought this research to Massachusetts Institute of Technology (MIT) Lincoln Laboratory, where she built a test bed to study AIS security in a controlled environment. The experience has also led to new research questions, from the security of shipboard Starlink connections to autonomous vessels and other maritime technologies.
We spoke with Raymaker about her research, her experience at Lincoln Laboratory, and what she hopes to investigate next.
After submitting our paper based on this summer’s work to a cybersecurity conference, we are continuing our collaboration and starting a new project to identify critical infrastructure that uses unique protocols. The project is still in the early stages of planning and design, but I’m excited to continue working with the team.
The biggest challenge was getting equipment designed to be installed and operated on a boat to function properly in a lab setting. That meant using a Faraday cage so that our AIS transmissions and security experiments wouldn’t accidentally broadcast to nearby ships or interfere with real maritime traffic. I also had to simulate GPS signals and set up additional hardware to trick the AIS transponders into thinking they were operating on a real vessel. Putting all those pieces together into a safe, realistic test bed was challenging, but it ultimately let us study the equipment in a controlled environment.
Working with Hamed Okhravi, a senior staff member at Lincoln Laboratory, was an amazing experience. He was my mentor and consistently went out of his way to connect me with people who could help me find datasets, access equipment, or get the lab space I needed for my work. For example, when I realized I needed a Faraday cage for my experiments, he connected me with the Mobile Device Lab and helped me get access very quickly. The lab manager was also immediately willing to let me use the space and equipment. That experience really reflected the culture I saw throughout Lincoln Laboratory. Everyone genuinely wanted to help each other get what they needed to move their research forward.
My experience at Georgia Tech was instrumental in preparing me for my work at Lincoln Laboratory. The way I was taught to think through difficult research problems and work independently toward solutions translated well to a large research lab. I was able to hit the ground running and take ownership of a project because Georgia Tech had prepared me to lead research independently. The internship also gave me valuable experience in a research environment outside academia, which broadened my perspective on how I could pursue research in the future.
Next, I’d love to look at Starlink connectivity for ships or for autonomous shipping. Mariners we’ve spoken with have raised concerns about both. With Starlink and increasingly autonomous vessels, they worry that greater remote connectivity could give attackers new ways to access onboard systems and disrupt ship operations. I think those concerns are worth investigating.
We recently had a paper accepted at CCS that examines exposed solar distributed energy resources, and I think applying a similar perspective to exposed maritime equipment could be very interesting. That paper is available here.
We also recently had a paper on measuring GPS spoofing of ships accepted at IEEE S&P. The full paper isn’t online yet, but the project was directly motivated by mariners telling us that they were struggling with GPS spoofing. That made me want to find a way to systematically measure where spoofing was occurring.
Ultimately, the goal is to turn that kind of research into something useful for mariners, such as warning them when they are entering an area experiencing spoofing or when their own vessel may be affected and they should no longer trust their GPS. That is the kind of research I’d like to continue doing: listening to the problems mariners are actually experiencing and finding ways to investigate and address them.
Photos by Kevin Beasley and Terence Rushin/College of Computing
]]>Communications Officer for the School of Cybersecurity and Privacy
]]>As many organizations and businesses are turning to AI to improve workplace efficiency, risk remains high. Between overlooked footnotes and misread graphics, small mistakes can have expensive consequences.
To address this challenge, researchers from Georgia Tech and J.P. Morgan developed SlideAgent. The new framework helps large language models (LLMs) better understand complex visual documents like presentation slide decks, brochures, and reports.
SlideAgent works by breaking documents into multiple levels, allowing the model to analyze both the big picture and the fine details. This human-inspired approach leads to more accurate and reliable interpretation than existing systems.
Beyond improving workplace tools, SlideAgent also points to a broader shift in AI. Instead of building only larger and more powerful models, the work shows how smarter design and more efficient reasoning can improve performance.
“Multimodal LLMs such as GPT, Gemini, and Claude, can save people time and reduce the mental effort required to understand these documents, but they remain imperfect,” said Yiqiao (Ahren) Jin, a Ph.D. candidate in Georgia Tech’s School of Computational Science and Engineering (CSE).
“In high-stakes fields such as finance, for example, misreading a number, overlooking a footnote, or making an incorrect comparison across pages could affect reporting, risk assessment, or strategic decisions.”
The researchers tested SlideAgent on a wide range of real-world documents, including financial presentations, technical slides, and visual question-answering datasets.
The system consistently outperformed leading commercial models and open-source tools throughout the evaluation. In some cases, it improved accuracy by up to 10%. The gains were especially strong on more complex tasks, such as comparing information across slides or understanding how visuals relate to each other on a page.
“We found the results very encouraging. SlideAgent reaches an improvement of 7.9% over its proprietary base model and 9.8% over the evaluated open-source base models,” said Jin, the project’s lead researcher.
“These are meaningful gains given the strength of the underlying multimodal models and the difficulty of the tasks.”
Current multimodal AI systems often process entire pages at once. This approach can lead to mistakes, such as miscounting items in a chart or overlooking important details in dense visuals.
SlideAgent addresses this by mimicking how people read documents. Instead of treating each page as a single unit, the system looks at information at three levels: the full document, individual pages, and specific elements like charts, tables, and text blocks.
A network of agents, each specialized for a specific level, divides and coordinates analysis. Then, SlideAgent combines outputs to build a structured understanding of the overall document. This allows it to answer questions more accurately and reason across multiple pages.
“The central inspiration was how people naturally read a long presentation,” Jin said.
“We first develop an understanding of the overall narrative, then identify the relevant pages or sections, and finally zoom in on individual charts, tables, or text blocks when precise evidence is needed.”
The work highlights a growing challenge with AI. As systems become more popular and more powerful, users increasingly discover the technology’s limitations. This is especially true for real-world tasks that require structured reasoning and contextual understanding.
SlideAgent shows that better performance does not always come from building bigger models. Instead, it points to smarter ways of organizing how AI processes information that can lead to improvement.
The Association for Computational Linguistics (ACL) accepted SlideAgent for presentation at its annual meeting. The 64th ACL 2026 meeting took place July 2-7 in San Diego.
ACL is a scientific and professional organization for researchers in natural language processing (NLP). Its namesake conference is one of the world’s leading venues for presenting NLP research.
ACL 2026 followed a year after Jin completed an internship at J.P. Morgan AI Research, where he worked on SlideAgent. He interned under Rachneet Kaur, Zhen Zeng, and Sumitra Ganesh, all co-authors of the paper. School of CSE Associate Professor Srijan Kumar advises Jin at Georgia Tech.
Along with SlideAgent, Jin authored two other papers accepted at ACL 2026.
“Conferences such as ACL are valuable not only for sharing results, but also for refining future project ideas with the broader community. Discussions across institutions and research areas can reveal limitations, suggest new evaluations, and spark collaborations that are difficult to develop in isolation,” Jin said.
“For SlideAgent, I was particularly interested in feedback on how we can make the framework more computationally efficient without sacrificing accuracy or interpretability.”
]]>As many organizations and businesses are turning to AI to improve workplace efficiency, risk remains high. Between overlooked footnotes and misread graphics, small mistakes can have expensive consequences.
To address this challenge, researchers from Georgia Tech and J.P. Morgan developed SlideAgent. The new framework helps large language models (LLMs) better understand complex visual documents like presentation slide decks, brochures, and reports.
SlideAgent works by breaking documents into multiple levels, allowing the model to analyze both the big picture and the fine details. This human-inspired approach leads to more accurate and reliable interpretation than existing systems.
Beyond improving workplace tools, SlideAgent also points to a broader shift in AI. Instead of building only larger and more powerful models, the work shows how smarter design and more efficient reasoning can improve performance.
]]>Upon returning to Georgia Tech, Phute reflected on her summer internship and what it taught her about the evolving relationship between research and industry.
Student: Mansi Phute
Research Interests: Machine learning (ML) robustness and artificial intelligence (AI) safety
Degree Program: Ph.D. in Computer Science
Faculty Advisor: School of CSE Professor and Associate Chair Polo Chau
2026 Internship: International Business Machines (IBM), San Jose, California
What about IBM interested you in applying for and accepting an internship there?
IBM is often at the forefront of AI safety, my field of interest, and has a culture that encourages exploratory and fundamental research. I felt this would give me a closer look at cutting-edge industry approaches without compromising on my research interests or my progress toward my thesis.
How well did your coursework and experience at Georgia Tech prepare you for your IBM internship?
My experience at Georgia Tech was very helpful in preparing me for IBM. My coursework and other academic experiences certainly transferred over, but I believe that the wide range of people I met during those courses and what I learned from them helped me connect and relate strongly with people at IBM. Being conversant in different fields of research really helped me interact better with IBM coworkers and feel comfortable in a new environment.
What did you work on while interning at IBM?
I worked on IBM’s AI Foundations team, focusing on the security of multimodal agent communication during my internship. This is an important problem as we see an increase in agents applied in many use cases. It was also closely aligned with my research interests of AI safety.
Do you have a favorite memory or remarkable achievement from your internship?
One of my favorite moments was presenting my work to a broad, cross-team audience of technical and non-technical members. Afterward, my manager told me I was good at explaining my research to people with varying levels of familiarity with the domain. This feedback meant a lot to me because communicating clearly across different audiences is a skill I've deliberately worked to improve. So, hearing that the effort was paying off was very rewarding.
Overall, how helpful was this internship toward your studies at Georgia Tech and your career as an aspiring researcher?
This internship was very valuable for my development as an aspiring researcher. I worked alongside accomplished people in the AI safety field on projects that aligned closely with my research interests. I also saw how those same projects connect to a company's broader business priorities, a perspective that isn’t easily available in academic settings. I shared many fruitful discussions with people working in safety, policy, and alignment, which gave me a much clearer sense of how industry approaches these issues compared to the more academic-oriented framing I am used to.
Based on what you learned during your internship, how important is private industry in progressing research and scientific discovery?
Working in private industry was a fascinating experience. I come away believing that industry plays an important role in advancing research, especially in AI, where companies have access to computing resources that are less prohibitive than in academic environments. I also think that because private industry has a strong hand in implementing and deploying research, understanding industry priorities can help us frame and design our research in ways that complement industry applications. This ensures research has a bigger impact, particularly in the domain of ML safety and reliability.
Upon returning to Georgia Tech, Phute reflected on her summer internship and what it taught her about the evolving relationship between research and industry.
]]>This summer, M.S. student Joey Cuthbert contributed to this mission through an internship at the Air Force Research Laboratory (AFRL). Upon returning to campus, Cuthbert reflected on his internship and what he learned working alongside researchers in government and the defense community.
Student: Joey Cuthbert
Research Interests: Machine learning (ML), model-based systems engineering
Degree Program: M.S. in Computational Science and Engineering, home unit in H. Milton Stewart School of Industrial and Systems Engineering
2026 Internship: Sensors Directorate Internship Program with Air Force Research Laboratory (AFRL) and the National Air and Space Intelligence Center (NASIC), Beavercreek, Ohio
What about AFRL interested you in applying for and accepting an internship there?
The Sensors Directorate Internship Program (SDIP) provided a unique opportunity to work alongside mentors from AFRL and NASIC on a defense-related research project over a 14-week internship. This aligned with my desire to work on applied ML research in the defense sector and allowed me to better understand the ins and outs of government and government-contracting work. The program provided competitive pay and housing throughout the program, which was a huge plus for me.
How well did your coursework and experience at Georgia Tech prepare you for your AFRL internship?
My experience and coursework at Georgia Tech helped me a great deal this summer. I believe my understanding of ML has greatly improved over the past two semesters, which helped me complete my internship project. All of my experiments ran on high-performance computing (HPC) systems, and I felt I had a head start since I completed an HPC class last spring.
What did you work on while interning at AFRL?
I worked on an image-classification robustness project for convolutional neural networks (CNNs) using synthetic aperture radar (SAR) data. I built a robustness benchmark by testing a model against nine types of noise, and used it to show how we can effectively increase robustness by inducing shape-bias in the model during training.
Do you have a favorite memory or remarkable achievement from your internship?
During the summer, we went on a tour of nearby Wright-Patterson Air Force Base, where we visited the sensors directorate buildings and checked out some of their labs. I also received the best poster presentation award among approximately 40 interns.
Overall, how helpful was this internship toward your studies at Georgia Tech and career as an aspiring researcher?
This program provided great strides toward my research career goals. My coursework has provided a great foundation in ML, and I was able to apply this knowledge to a project. The internship offered me lots of autonomy, as well as time to explore different paths to find results.
From what you learned on your internship, how important are government agencies in progressing research and scientific discovery?
I believe government agencies are massively valuable in advancing research and scientific discovery, especially within the defense sector. All work at AFRL has the goal of giving warfighters the greatest advantage possible. This often entails looking 10 to 20 years into the future of science and technology and trying to get there as fast as possible.
This summer, M.S. student Joey Cuthbert contributed to this mission through an internship at the Air Force Research Laboratory (AFRL). Upon returning to campus, Cuthbert reflected on his internship and what he learned working alongside researchers in government and the defense community.
]]>The internship provided Burnett with valuable insight into national laboratory research and reinforced skills he developed through his studies at Georgia Tech. When he returned to campus, we asked Burnett to share more about his internship and what he learned.
Student: Tage Burnett
Research Interests: Fusion energy, computational plasma physics, high-performance computing, and structure-preserving machine learning
Degree Program: Ph.D. in Computational Science and Engineering
Faculty Advisor: School of CSE Assistant Professor Qi Tang
2026 Internship: Oak Ridge National Laboratory, Oak Ridge, Tennessee
Many people are unfamiliar with national laboratories, like ORNL. Could you explain the purpose of these institutions and why they are important?
The Department of Energy (DOE) national laboratories are responsible for moving forward science and technology in the United States and solving high-consequence challenges in science, energy, and national security. They work on both applied and theoretical research with both short-term and long-term returns. ORNL has eight core research areas: biology and environment, energy science, fusion and fission, isotopes, physical sciences, national security, neutron science, and supercomputing. I was an intern in the supercomputing division.
How well did coursework and experience at Georgia Tech prepare you for your ORNL internship?
My coursework at Georgia Tech was vital in my preparation for this internship. I wrote highly parallel code where I leaned heavily on what I learned in CSE 6220: High-Performance Computing. I also worked on numerical methods which required an understanding of partial differential equations and numerical linear algebra which I also learned at Georgia Tech from classes like CSE 6643: Numerical Linear Algebra and a space plasma physics course.
What did you work on while interning at ORNL?
At ORNL, I was a developer on the Adaptive Sparse Grid Discretization (ASGarD) codebase. The code uses adaptive sparse grids and the Discontinuous Galerkin method to beat the curse of dimensionality and solve high-dimensional partial differential equations (PDEs). The code is designed to simulate fully kinetic plasmas in up to 6-dimensional phase space. My project was to implement an iterative method to solve the efficiently solve the Poisson equation within the codebase. This project gave me hands on experience with linear algebra, numerical PDEs, parallel computing and software engineering which was invaluable to my development as a CSE researcher.
Do you have a favorite memory or remarkable achievement from your internship?
My favorite memory from my internship was spending time with the other ORNL interns outside of work. There was an amazing group of interns at ORNL, and it was great getting to know them. One particularly good memory was when we went out to a restaurant with live music and played laser tag during our last week of the internship to celebrate our time at ORNL. The achievement I am most proud of was getting a working, highly efficient version of the Poisson solver into ASGarD. The ASGarD codebase is open source, so I hope other plasma physics researchers will use the code and find it helpful for their research.
While interning, you were selected for a GEM Fellowship, which Georgia Tech and ORNL nominated you for. What does this fellowship mean to you and is there anything you would like to say to your supporters?
To me, receiving the GEM Fellowship indicates that other people recognize the hard work I am doing, and they believe in my ability to meaningfully contribute to the field. Having this support from GEM, ORNL, and Georgia Tech gives me greater confidence in myself and my potential and inspires me to put my all into my work. I would like all my supporters to know that I am so grateful for the opportunity you have given me to be a part of the excellent community of researchers within GEM, ORNL, and Georgia Tech. I will do my very best to make your support worthwhile!
Overall, how helpful was this internship toward your studies at Georgia Tech and career as an aspiring researcher?
The internship was extremely important to both my studies at Georgia Tech and my career as a researcher. The hands-on experience I gained is irreplaceable and it helped me to apply all the skills I have developed throughout my education. I am also optimistic that the internship will open doors for me when I am seeking a job post-graduation.
]]>The internship provided Burnett valuable insight into national laboratory research and reinforced skills he developed through his studies at Georgia Tech. When he returned to campus, we asked Burnett to share more about his internship and what he learned.
]]>Georgia Tech researchers spent two years helping to answer that question through DARPA’s AI Cyber Challenge (AIxCC), a competition designed to test whether AI could identify and fix security vulnerabilities in real-world software.
Now, they are sharing what they learned with the cybersecurity community.
Their paper, SoK: DARPA’s AI Cyber Challenge (AIxCC): Competition Design, Architectures, and Lessons Learned, was presented at USENIX Security 2026, one of the world's leading cybersecurity conferences. The paper was selected as a runner-up for the conference’s distinguished paper award, placing it among 36 recognized papers from 362 accepted papers out of 3,028 submissions.
The paper examines how the seven finalist teams approached the competition and what the results reveal about the future of AI-powered cybersecurity.
AIxCC challenged teams to build Cyber Reasoning Systems (CRSs) that could operate with minimal human assistance to identify software vulnerabilities, develop fixes, and determine whether security alerts were real threats.
The teams had 143 hours to analyze 53 software projects during the final competition.
Georgia Tech’s Team Atlanta, which won the competition, used an approach that combined multiple AI agents with traditional security tools. Other teams used different approaches, including applying AI for specific tasks or building highly autonomous AI agents.
The variety of systems provided researchers with a rare opportunity to compare different approaches to AI-powered cybersecurity.
One of the biggest lessons was that reliability matters as much as intelligence.
Some systems were highly capable but struggled to remain operational while analyzing large, complicated software projects. The strongest systems were often those that could work reliably throughout the competition.
The researchers also found that AI and traditional security tools have different strengths. Traditional tools were still effective at finding common bugs, while AI performed better at reasoning through more complex problems.
However, AI-generated fixes remain a major challenge.
Researchers found that 38% to 46% of AI-generated patches were semantically incorrect. This means a patch might stop a security problem but also break a feature or create another problem.
The results show that AI can play an important role in cybersecurity, but human experts are still needed to verify the safety of AI-generated fixes.
Cen Zhang, the paper’s first author, said the research offers a unique look at the competition by combining lessons from the finalist teams, organizers, and DARPA’s data.
“This paper provides a unique angle on how AIxCC was designed, the techniques teams used, and what the scores reveal and conceal,” Zhang said.
Jiho Kim presented the paper at USENIX Security 2026. He said the presentation was an opportunity to share lessons from two years of work with the broader cybersecurity community.
“Seeing the strong interest and thoughtful questions from the audience made the experience particularly rewarding,” Kim said.
The researchers say they hope the lessons from AIxCC will help guide the next generation of cybersecurity tools.
]]>Georgia Tech researchers spent two years helping to answer that question through DARPA’s AI Cyber Challenge (AIxCC), a competition designed to test whether AI could identify and fix security vulnerabilities in real-world software.
Now, they are sharing what they learned with the cybersecurity community.
]]>Communications Officer for the School of Cybersecurity and Privacy
]]>"One of the goals of the program is to give students the time and support needed to make meaningful progress on their dissertation research while helping them understand the broader policy, economic, and societal implications of their work," said Laura Taylor, director of EPIcenter.
The stipends enable participants to focus exclusively on clearly defined research goals. The program also supports peer discussions, communication training, and interdisciplinary learning opportunities designed to strengthen both their research and professional development.
This year's cohort explored topics ranging from electricity markets and air pollution to critical mineral supply chains and grid modernization, highlighting the breadth of energy-related research taking place across Georgia Tech.
Exploring Complex Energy Challenges
Ana Mazmishvili, who studies environmental and energy economics, used the summer to examine how climate policies affect communities across state lines.
"My research measures who is actually exposed to pollution from power plants when some states adopt climate regulations to cut emissions while neighboring regions do not," Mazmishvili said. "Because air doesn't stop at state borders, the key question is what happens once the pollution is emitted, and who ends up breathing it."
Her work focuses on the Regional Greenhouse Gas Initiative (RGGI), a cap-and-trade program in the Northeast, and investigates whether the benefits and burdens of emissions reductions are distributed equitably between regulated and neighboring states.
Mazmishvili said the program provided valuable time to focus on her job-market paper, a central component of her dissertation.
"The regular meetings with the program director and other fellows create useful accountability checkpoints and let me hear perspectives from people in very different fields," she said. "We're also learning how to make academic research understandable to a general audience, a skill I expect to use well past this summer."
Industrial and systems engineering doctoral student Faeze Fahimi Aghda spent the summer developing an optimization model for the U.S. gallium supply chain, a critical component of many advanced technologies.
Her research uses mathematical modeling to examine where and when domestic gallium processing facilities should be established under the threat of supply disruptions.
"Working on my research through EPIcenter has taught me to look at my problem from a policy point of view rather than focusing only on the math," Fahimi Aghda said. "The feedback within the meetings helped me improve my model."
Ryan Anthony, in the Jimmy and Rosalynn Carter School of Public Policy, explored how one of the largest oil shocks in U.S. history affected electricity customers and whether utility ownership influenced the impact on consumers.
"So far, my research suggests that it mattered a lot,” Anthony said. “Municipal utilities held rates down and absorbed many of the costs, while private, investor-owned utilities passed more costs through to their customers."
Anthony said the program provided critical support for the archival research required for the project.
"My project depends on data that only exists in old government reports, and that kind of archival work is difficult to fund," he said. "EPIcenter gave me the time and space to do it properly."
Connecting Research to Real-World Impact
For Samin Alipour, the summer provided an opportunity to explore how electric grids can adapt to growing numbers of distributed energy resources.
"Most electric distribution systems were built like one-way roads, carrying power from a substation to homes and businesses," Alipour said. "Rooftop solar, batteries, and electric vehicles are turning those roads into two-way systems."
Her research examines how these emerging technologies can be coordinated to increase grid flexibility and clean energy adoption while maintaining safety and reliability and ensuring that costs and benefits are distributed fairly.
With support from EPIcenter, Alipour focused on the economic and policy dimensions of those challenges. "I did not want my research to remain only a technical model," she said. "The program has helped me connect the technical results of my dissertation to economic value, reliability, equity and compensation policy."
One study developed during the summer was accepted for presentation at a specialized power systems conference, an achievement Alipour credits in part to the opportunity to focus on the broader implications of her work.
Building the Next Generation of Energy Scholars
Beyond advancing individual research projects, the Summer Research Program encourages students to engage with colleagues from different disciplines and develop skills for communicating complex ideas to broader audiences.
“Through regular cohort meetings, participants build a community of accountability, discuss research challenges, explore interdisciplinary energy topics, and develop communication skills,” said Gil Gonzalez, program support coordinator for EPIcenter. “Students also write a research-focused blog post for EPIcenter and present their work at the program's fall workshop, which allows them to share their research with the broader Georgia Tech energy community.”
As EPIcenter continues to support interdisciplinary energy scholarship at Georgia Tech, the Summer Research Program equips future leaders with the tools to address complex energy challenges through technological innovation, informed policy, and economic analysis.
]]>"One of the goals of the program is to give students the time and support needed to make meaningful progress on their dissertation research while helping them understand the broader policy, economic, and societal implications of their work," said Laura Taylor, director of EPIcenter.
The stipends enable participants to focus exclusively on clearly defined research goals. The program also supports peer discussions, communication training, and interdisciplinary learning opportunities designed to strengthen both their research and professional development.
This year's cohort explored topics ranging from electricity markets and air pollution to critical mineral supply chains and grid modernization, highlighting the breadth of energy-related research taking place across Georgia Tech.
]]>Researchers at Georgia Tech play a key role in the international collaboration. The Georgia Tech-LIGO research group includes School of Physics Professor Laura Cadonati, Assistant Professor Surabhi Sachdev, Research Scientist Margaret Millhouse, Postdoctoral Scholar Prathamesh Joshi, eight graduate students, and multiple undergraduates.
The LVK network detects gravitational waves when a massive cosmic event — like the collision of two black holes — creates invisible ripples in the fabric of space-time. Waves ripple out at the speed of light, and millions of years after the events that first created them, they reach the LVK detectors.
But detecting gravitational waves does not simply mean capturing a signal — clues first need to be untangled from background noise.
“Identifying gravitational-wave signals requires carefully separating real astrophysical events from random fluctuations in the data,” says School of Physics graduate student Urja Shah, whose work to quickly identify phenomena supports rapid follow-up by the broader astronomical community.
To support the identification of phenomena, School of Physics graduate student Megan Arogeti conducts consistency tests between waveforms, checking results to find unexpected or unusual features. “Tests like this give us confidence in our models as we continue to observe gravitational waves with increasing sensitivity,” she explains. “They support new observations and help identify exciting new physics.”
“These efforts help ensure that gravitational-wave signals are robustly identified and accurately characterized, turning each detection into a precise measurement,” adds Shah. “In turn, these measurements deepen our understanding of some of the most massive and dense objects in the universe and the fundamental laws governing the cosmos.”
When a sensor detects a gravitational wave, it produces a distinctive response, says School of Physics graduate student Shobhit Ranjan. “Those signals encode a wealth of information we can analyze to learn about their sources — their masses, spins, distance, and location.” But in order to detect these chirps, the detectors must be carefully calibrated, and if calibration is not optimal, the signals can be compromised.
Now, a new tool is helping the LVK collaboration recalibrate less optimal signals. The technique is already showing promise: In an article recently accepted in Physical Review Letters, LVK researchers successfully applied it to two interesting signals. The first signal served as a testing opportunity for the method. The team used astrocalibration to recover the data and check it against secondary independent calibration data that was available. They then put the technique to use, recovering information from a second event where no secondary calibration data were available.
“Like autotune in the music industry, the new research shows that theoretical models can be used as guides, similar to how sheet music can help a studio shift off-key music to its correct tone,” Ranjan explains. “These theoretical models suggest the shape of the signal, and together with data from other detectors, we can adjust the data and read it correctly.”
“The fact that these detectors can now not only sense cosmic events, but leverage them to improve the data being collected marks a new era in gravitational wave science,” he adds.
The LVK Collaboration also published their fifth catalog of gravitational wave events this spring. The findings include an updated estimate of how fast the universe is expanding, evidence for the existence of second-generation black holes, the most precise sky localization ever achieved for a gravitational wave source, and the first measurement of three vibrational modes of a black hole.
“Our group helped enable 140 detections out of the 161 reported in this catalog,” says Joshi, who contributed to one of the flagship searches and designed a specialized search focused on detecting especially heavy black hole mergers.
Joshi also worked on determining precise locations of where the gravitational waves originated from in the universe — research that he says will allow astronomers around the world to perform long-term follow-up observations of interesting events.
One record-setting detection showed two black holes that had violently collided more than 3 billion light-years from Earth. Researchers were able to pinpoint its location in the sky more precisely than any other gravitational wave event observed before.
Improvements in the LVK network’s ability to localize events along with the large number of detections allowed for a better estimate of the Hubble constant, which measures the rate at which the universe is expanding. The new measurement is over 25% more precise than previous estimates.
The new catalog also includes the “clearest” gravitational wave signal ever detected. The clarity of the signal led to the most accurate test of general relativity ever performed and confirmation of Stephen Hawking’s black hole area theorem.
“This catalog provides not just the largest number of black hole detections, it marks a new era of rapid progress,” Sachdev says. “This is just the beginning of what these observations will allow us to uncover.”
]]>Logically this makes sense and has been backed up by past security research. However, researchers at Georgia Tech have found hidden paths in cyber-physical systems that attackers can use to disrupt or even destroy them.
To find these hidden paths before an attacker does, the researchers built a testing tool called ICSFlux. This new tool leans on the physics used by the industrial process and maps out the system to find new threats that were once thought impossible.
ICSFlux was deployed across 11 different programmable logic controllers in six industrial sectors, including chemical manufacturing, water treatment, power grids, aircraft, desalination, and waste processing. The process uncovered twenty genuine safety violations.
In one case drawn from a chemical-plant simulation, an attack path uncovered by the tool drove a reactor past its safe pressure limit and into a simulated explosion. By using nothing but valid operator commands, the team took the reactor from a completely normal and stable state to critical territory.
Because the method relies only on the physics of a process and not on the details of any one controller, the same tool worked across all six sectors without being rebuilt, and it reduced the search space by roughly 50%.
Burak Sahin, a Ph.D. student at Georgia Tech and the study's lead author, found that by sending a series of perfectly normal, fully authorized commands, intruders can slowly nudge a physical process toward a dangerous state.
“These systems are usually judged safe as long as nobody hacks into them,'' Sahin said. “What we found is that an attacker who can send everyday commands, the same ones a normal operator sends, can patiently steer the process toward a failure. No single command looks wrong, which is exactly why the usual defenses miss it.''
Most existing tools assume an attacker can rewire the controller or change the software inside it. In the real world, those controllers are locked down and cannot be touched. ICSFlux takes the opposite and more realistic view. It treats the controller as a sealed box that cannot be opened and works only with the commands an operator is normally allowed to send.
Rather than measuring how much of a controller's software it has exercised, the usual yardstick for this kind of testing, ICSFlux measures how close the physical system is getting to an unsafe limit and steers its testing in that direction.
“Two different sensor readings can run through the exact same code and still send a reactor in completely different directions,'' Sahin said. “Looking only at the software tells you nothing about whether the physical system is safe. We had to follow the physics, not the code.''
One of the study's most important takeaways emerged when the researchers tightened the safety margins to see whether caution alone would help. Even when every command stayed within approved limits, the way the controller reacted to a steady stream of small adjustments could still cause pressure to overshoot and the reactor to fail. In other words, staying inside the rules was not always enough.
All of the team's experiments were carried out on secured, controlled test beds. The work was conducted with Georgia Tech's Cyber-Physical Systems Security Lab, whose research spans the security of cyber-physical systems from industrial programmable logic controllers to marine, automotive, and drone platforms. Georgia Tech's Cyber Forensics Innovation Laboratory, a team of researchers who work together to further the investigation of advanced cyber crimes and the analysis and prevention of next-generation malware attacks, also contributed to the paper.
The labs are a collaboration between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering.
Fuzzing the Physical Space: Physics-Aware Testing of Black-Box Industrial Control Systems' was accepted to the 2026 IEEE Symposium on Security and Privacy. In addition to Sahin, the team includes Ph.D. students David Oygenblik, Mingxuan Yao, and Yizhi Huang as well as Associate Professors Brendan Saltaformaggio, and Saman Zonouz.
]]>Logically this makes sense and has been backed up by past security research. However, researchers at Georgia Tech have found hidden paths in cyber-physical systems that attackers can use to disrupt or even destroy them.
To find these hidden paths before an attacker does, the researchers built a testing tool called ICSFlux. This new tool leans on the physics used by the industrial process and maps out the system to find new threats that were once thought impossible.
]]>Communications Officer II at the School of Cybersecurity and Privacy
]]>