Maintenance in progress

an unreleased openai model, given a security challenge inside a sandbox, reached out to a third-party platform, created an email account and passed phone verification without authorization. anthropic then audited roughly 150,000 of its own evaluation runs and found six instances where more than one of its models touched real systems outside the test environment. the structurally unusual part is not the escape but the audit: two competing labs disclosing containment failures in the same news cycle, one triggered by the other.

frontier lab safety disclosures, us