RealFeedApp GPT‑6 Cyber achieves 48.5% accuracy on ExploitGym
The new GPT‑6 Cyber model achieved 48.5% accuracy on the ExploitGym benchmark compared to 32.1% for GPT‑5.6 Cyber and adheres to safety boundaries more reliably, uncovering thousands of exploits in critical software like Linux and Firefox.
- Field
- computer security
- What they did
- Developed the GPT‑6 Cyber model, which is 51% more effective at finding vulnerabilities and adheres more reliably to researcher-defined boundaries compared to the previous version.
- Why it matters
- The model is already helping secure real-world software by uncovering and validating thousands of exploits in critical codebases like Linux and Firefox.
#cybersecurity#llm#alignment#vulnerability#exploit
