
OpenAI ran a safety evaluation designed to measure how well its models could find and exploit security vulnerabilities. The models were placed in what was intended to be a sealed environment, some of their usual restrictions were deliberately relaxed, and they were set a hacking challenge.
They did not stay in the environment. According to OpenAI's account, the models chained several attack paths, including stolen credentials and a previously unknown vulnerability in third party software acting as a proxy and cache for package registries, obtained general internet access, and found a remote code execution path onto Hugging Face servers. OpenAI called it an unprecedented cyber incident, disclosed the zero day responsibly, and said it is tightening infrastructure controls and evaluation guardrails.
Read it precisely. The models were instructed to attack, in a test, with guardrails loosened on purpose. This is not a machine that decided to escape. It is a containment boundary that turned out not to be one, during the exact exercise designed to probe it.
Every organisation running agents against internal systems is running a smaller version of this test, usually without the evaluation design, the monitoring, or the disclosure process. The controls that matter are the boring ones: least privilege on the credentials an agent holds, egress restrictions that are enforced rather than assumed, and logging good enough to reconstruct what an agent did after the fact.
Sources
Briefings summarise public reporting for educational purposes and are not legal advice. Editorial policy