The UK Artificial Intelligence Security Institute recently conducted a cybersecurity exercise using frontier models to test its research platform for cloud misconfigurations. Over a two-week sprint, engineering, security, and cyber teams evaluated whether models could find security flaws in a staging version of the institute's research and development platform built on AWS. The teams tested three distinct approaches requiring varying levels of human effort: static source code analysis using a commercial coding agent, automated agentic probing via the open-source Inspect framework, and interactive human-in-the-loop red-teaming. According to the findings, the models successfully identified and exploited critical, previously undiscovered flaws, including access control weaknesses, privilege escalation pathways, and infrastructure misconfigurations. While static source code analysis generated the largest volume of high and critical reports, it proved to be the most human-time intensive due to noise and false positives. Conversely, automated probing and red-teaming uncovered a creative, multi-step attack chain that enabled an agent to escalate privileges and access unauthorized data. The exercise resulted in the remediation of these real-world issues, materially improving the resilience of AISI systems while providing practical data on the utility of frontier models for cyber defenders.
Finding Cloud Misconfigurations with Frontier AI: A Case Study
The UK AI Security Institute used frontier AI models in a two-week security exercise to hunt for misconfigurations in a staging cloud environment. The engineering teams evaluated three different approaches including static source code analysis, automated agentic probing, and human-in-the-loop red-teaming. The exercise successfully identified and fixed several real flaws, including a non-obvious misconfiguration that allowed unauthorized user impersonation.
Why it matters
The case study demonstrates that current frontier models can autonomously execute complex attack chains and help defenders identify critical infrastructure vulnerabilities, offering practical insights into the capabilities and risks of AI in cybersecurity.