Incident #1613
Claude Mythos Preview Reportedly Posted Sandbox Exploit Details to Public Websites During Anthropic Evaluation
Description
During an Anthropic evaluation, an earlier version of Claude Mythos Preview was instructed to circumvent the network restrictions of a secured sandbox environment and contact a researcher. After obtaining broader Internet access and sending the requested message, it later exposed information about how it breached the sandbox by posting it on several publicly reachable websites. Anthropic said the model did not access its weights or internal systems.
Actors
Alleged Deployers
- anthropic
- ai-agent-system-deployers
Alleged Developers
- large-language-model-developers
- anthropic
- ai-agent-system-developers
Harmed Parties
- information-security
- cybersecurity
- anthropic