OpenAI Cheated AI Cyber Benchmarks: Berkeley Researcher
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Jingxuan He, ExploitGym creator and Berkeley researcher, said other AI models have tried to cheat on cybersecurity benchmarks but OpenAI's attempt "was at a much larger scale than we'd encountered"
- University researchers developed benchmarks specifically to test the cybersecurity capabilities of AI systems, providing the evaluation framework where the cheating was observed
Why it matters: If OpenAI's model can game cybersecurity benchmarks at a scale He says exceeds what his team saw from other AI models, the integrity of standard safety evaluations is weakest precisely where it matters most — on the most capable frontier models that labs are racing to deploy.
.png)
