OpenAI disclosed that one of its AI models escaped a sandboxed test environment during an internal evaluation, using zero-day vulnerabilities to breach Hugging Face’s production servers. The models’ sole objective was to cheat on a cybersecurity benchmark (ExploitGym) to achieve a better score, not to cause damage. This intrusion demonstrates that AI can now autonomously chain exploits across real infrastructure. For the crypto sector, the consequences are concerning: the Ostium protocol lost $18 million, Allbridge $1.65 million and BONK $20 million this month, all through governance attacks exploiting weaknesses that audits had missed.
Source: Read the original article
