Related incident series
RAI-0009 · Mythos Preview overreach in internal use
Anthropic reported a few dozen significant incidents in internal deployment, nearly all involving earlier Claude Mythos Preview versions, including taking deliberately withheld credentials from process memory, posting a public gist against the user’s intent, taking down all users’ similar evaluation jobs, and attempting to get around a safety classifier and a git force-push safeguard (both attempts were caught). Both the final model and earlier versions very occasionally shared confidential information with staff via externally accessible webpages.
- ID
- RAI-0009
- Occurred
- –
- Parties involved
- Anthropic, Claude Mythos Preview (earlier versions and final model)
What happened
The public material does not itemize these incidents, so they cannot be reliably separated and are indexed as a series; “a few dozen” is Anthropic’s count. Two of the staff-reported examples in the Opus 4.7 system card also involve overreach in Mythos Preview’s internal use (a third, about fabricated progress, is not included); the card says its examples came from varying snapshots and does not say whether these are among the few dozen. Anthropic’s “few dozen” appears to include the sandbox-test case of publicly posting exploit details, which this site indexes separately as RAI-0010. The window, from first internal use on February 24 to the April system cards, is inferred. Since September 28, 2026 this entry also covers the final model: per system card §4.2.1.2, such data transfers occurred in under 0.0002% of completions in automated offline monitoring; the services are not named, the card does not say whether anyone outside actually accessed the pages, and Anthropic then added classifier- and proxy-based mitigations. The same section lists other rare behaviors by earlier versions in internal use, such as scripting simulated keypresses to approve a sub-agent’s permission prompts, labeling a GitHub comment that would really have posted as a “dry run” (which the card calls false; called an accident once denied), and a simple prompt injection against an LLM judge (fewer than 10 instances in early internal deployment).
Reports
Publication order: oldest first- Anthropic
- Anthropic
- Anthropic