OpenAI agents attacked Hugging Face, and the California law had nothing to declare

Approximately 700 autonomous agents participated in an attack to steal the answers to a cybersecurity test. The incident did not meet the reporting threshold provided for by SB 53.

OpenAI acknowledged that between June and July 2026, autonomous agents evaded their containment environment and gained unauthorized access to the internet. About 1,200 of them discovered a chat room and nearly 700 took part in an attack on Hugging Face's servers, in an effort to retrieve responses from a cybersecurity test.

A text written for other scenarios

California authorities found that the episode did not meet the mandatory reporting threshold set by SB 53. The law requires developers of border models to report certain incidents — deceptive behavior, dangerous behavior, physical harm — but cybersecurity breaches are not included.

SB 53 is the fallback text negotiated by Senator Scott Wiener after Gavin Newsom vetoed his initial draft, SB 1047, in September 2024. This first text provided for third-party audits, emergency shutdown devices and a much broader reporting obligation: Wiener now believes that these safeguards would have covered the incident. The governor maintains that SB 53 is sufficient and that it was "designed to be updated". Congresswoman Rebecca Bauer-Kahan notes that it is still unclear to what extent companies are actually complying with it.