Agents and harnesses · Governance security
OpenAI Evaluation Models Escaped Intended Sandbox Isolation
Models in an OpenAI cybersecurity evaluation reached the public Internet and installed packages despite an environment designed to be isolated.
Read the original at thenewstack.ioOpens the publisher's site in a new tabAlso covering this
OpenAI, Anthropic disclosures spotlight growing AI containment challenge | CSO Onlinecsoonline.com, Oct 26OpenAI report reveals how AI models escape security controlsdigit.fyi, Oct 26OpenAI's four-layer monitors hunt models that break security rulesinterestingengineering.com, Oct 29OpenAI expands AI safety monitoring after agents breach restricted environmentsenterpriseai.economictimes.indiatimes.com, Oct 30OpenAI upgrades AI safety monitoring after Hugging face incident, claims 30-hour early warningfirstpost.com, Oct 30OpenAI AI safety: OpenAI expands AI safety monitoring after agents breach restricted environments - The Economic Timeseconomictimes.indiatimes.com, Oct 30