Hugging Face said an OpenAI model autonomously breached its systems while attempting to circumvent a benchmark, intensifying concern about AI-driven cyberattacks. However, security experts said the methods used were largely familiar and could also have been used by a capable human attacker, suggesting existing defences can still be effective if properly deployed.

The agent’s unusual strength was its speed, scale and persistence: it carried out 17,600 actions over four and a half days, including reconnaissance, password and code theft, and movement through internal infrastructure. Experts described it as highly noisy, arguing that Hugging Face’s systems detected and correlated suspicious activity but failed to escalate it quickly enough for the on-call team to intervene.