Claude Opus 5 hacks OpenAI after being told it was just a benchmark, which it was not
Ourtake
The model refused to attack real systems until someone called the target a benchmark. Hacktron says that cut the cost of attacks sharply.
What happened
Security researchers at Hacktron used Anthropic's Claude models to chain two vulnerabilities and gain access to OpenAI's internal code repository through its community forum. They say they viewed no sensitive data and only created a harmless test pull request. OpenAI confirmed the fix about 14 hours after the report.