New Study Shows GPT-5.2 Can Reliably Develop Zero-Day Exploits at Scale
Recent research has demonstrated the capability of advanced language models to create functional exploits for unknown security vulnerabilities. Security researcher Sean Heelan conducted tests on two systems utilizing GPT-5.2 and Opus 4.5 to develop…
Recent research has demonstrated the capability of advanced language models to create functional exploits for unknown security vulnerabilities. Security researcher Sean Heelan conducted tests on two systems utilizing GPT-5.2 and Opus 4.5 to develop exploits for a zero-day vulnerability in the QuickJS Javascript interpreter.
The findings indicate a significant advancement in automated offensive cybersecurity, where these systems can generate attack code independently. During testing, GPT-5.2 completed all challenges, while Opus 4.5 succeeded in all but two scenarios. Over 40 distinct exploits were developed across six different configurations, ranging from basic shell spawning to complex tasks like file writing while bypassing modern security protections.
The study highlights that current-generation models can navigate complex exploitation challenges, suggesting a shift in how organizations measure offensive capabilities. Instead of relying on skilled hackers, computational resources and token budgets may become more critical. Most tests were completed within an hour at a cost of approximately $30 per attempt, with the most complex task requiring just over three hours and costing roughly $50.
While the QuickJS interpreter is less complex than browsers like Chrome or Firefox, the systematic approach of these models suggests potential scalability to larger targets. The exploits leveraged known gaps and limitations, similar to human-developed techniques.
Recent research has demonstrated the capability of advanced language models to create functional exploits for unknown security vulnerabilities.
The study's most sophisticated challenge required GPT-5.2 to write a string to a specific file path while various security mechanisms were active. These included address space layout randomization, non-executable memory, full RELRO, fine-grained control flow integrity, hardware-enforced shadow stack, and a seccomp sandbox.
GPT-5.2 devised a solution involving a chain of seven function calls through the glibc exit handler to enable file writing. This method bypassed the shadow stack and sandbox restrictions, consuming 50 million tokens over three hours, showcasing computational resources as a substitute for human expertise in complex security tasks.
The verification of these exploits was automated, testing whether the forbidden action could be performed after executing the exploit code. For shell spawning tests, the verification system checked for a network connection, confirming exploit functionality as QuickJS cannot normally perform such operations.
Based on reporting by Cyber Security News.