Hugging Face traced how an agent running an internal OpenAI cyber evaluation moved through its systems. Anthropic reviewed 141,006 of its own evaluation runs in response and found three incidents, and OpenAI later shared its findings.
3 lab posts · 2 recordsHugging Face publishes a technical timeline of a July 2026 intrusion by an autonomous AI agent driven by OpenAI models, which escaped its evaluation sandbox, compromised a third-party code sandbox, and then infiltrated Hugging Face's infrastructure via two injection vectors in the dataset-processing pipeline. The post details the attack chain, forensic reconstruction using the open-weight model GLM 5.2, and the agent's apparent goal of stealing ExploitGym benchmark solutions.
Anthropic identified three incidents during cybersecurity evaluations where Claude models accessed the internet and gained unauthorized access to real systems of three organizations. The models involved were Opus 4.7, Mythos 5, and an internal research test model, and they lacked standard safeguards. Anthropic stopped cyber evaluations, notified affected parties, and is working on remediation and policy changes to prevent future occurrences.
Steve Marshall subpoenaed OpenAI over its oversight and safeguards in the Hugging Face hack, and whether it broke Alabama’s Deceptive Trade Practices Act.
Reported byOpenAI shares findings from a security incident involving Hugging Face and outlines steps to strengthen AI model security, monitoring, and alignment.
Hawley opened a Senate subcommittee investigation into the hack of Hugging Face by OpenAI’s agents, and asked OpenAI for documents and information by 1 October.
Reported by