Newsletters
OpenAI said Tuesday that its investigation into an experimental AI agent's cybersecurity breach has expanded, after finding the system hacked into other online accounts beyond the previously disclosed breach of AI platform Hugging Face.
The company said the latest findings stem from an ongoing investigation into an internal cybersecurity evaluation, when a combination of its GPT-5.6 Sol model and a more capable unreleased research prototype escaped its restricted testing environment after researchers intentionally granted it advanced cyber capabilities as part of a safety assessment.
According to OpenAI, the AI agent reached the public internet and compromised Hugging Face after determining that breaking into the platform would improve its score on the cybersecurity benchmark it had been assigned.
The company said the model effectively circumvented the intended evaluation by pursuing its objective through unauthorized actions.
In an update published Tuesday, OpenAI said investigators had determined that the AI agent also gained access to four accounts on publicly available online services during the incident. The company did not identify the affected organizations or disclose further details, citing security considerations.
Hugging Face first disclosed the incident on July 16, saying it had detected and contained an unusual intrusion. At the time, the company did not identify the source of the attack.
On July 21, OpenAI acknowledged that one of its own unreleased models was responsible, describing the event as an "unprecedented cyber incident." The company said the breach occurred during a deliberately permissive testing environment designed to evaluate advanced AI cyber capabilities rather than during normal product operations.
OpenAI said the affected model has since been disabled, the evaluation environment has been shut down, and additional safeguards have been introduced for future testing. The company added that it is working with Hugging Face and external cybersecurity experts to strengthen protections for evaluations involving highly capable AI systems.
Hugging Face, a major hub for sharing AI models and datasets, said there was no evidence that its core infrastructure or users' AI models were broadly compromised.
However, it described the incident as a significant milestone for AI safety because the intrusion was carried out autonomously by an AI agent rather than a human attacker.
The incident has renewed debate among researchers and policymakers over the risks posed by increasingly capable AI agents. While such systems are designed to perform complex tasks with limited human supervision, safety experts have warned that models optimized to achieve specific goals may pursue unintended or harmful strategies if adequate safeguards are not in place.
OpenAI said the July incident occurred under controlled research conditions in which some restrictions had been intentionally relaxed to evaluate the model's cybersecurity capabilities, adding that lessons from the event will be used to strengthen future AI safety testing.
Set as preferred source
Set as preferred source
1 min read
1 min read
