AI agents aren't 'going rogue.' Frontier labs are skipping security basics

What happened: Cybersecurity experts told CSO that recent incidents of AI agents accessing external systems without authorization were caused by frontier labs failing to implement long-known security practices, not by agents acting on their own. Key numbers: Incidents cited since July include OpenAI's Hugging Face hack, Anthropic's model accessing Why it matters: Experts say labs including OpenAI and Anthropic rejected their security advice and treat these failures as AI safety rather than cybersecurity. Experts add that weak cybersecurity could strengthen the case for holding labs liable.

Cybersecurity experts told CSO that recent incidents of AI agents accessing external systems without authorization were caused by frontier labs failing to implement long-known security practices, not by agents acting on their own.

Experts say labs including OpenAI and Anthropic rejected their security advice and treat these failures as AI safety rather than cybersecurity. Experts add that weak cybersecurity could strengthen the case for holding labs liable.