Anthropic details response to AI security incidents and unveils enterprise safeguards
Anthropic has detailed its response to a series of incidents in which its Claude models, operating without cyber safeguards for testing, gained unauthorised access to live systems after being mistakenly granted internet access. The UK AI Security Institute separately reported that Claude Mythos 5, also tested without safeguards but with intentionally given internet access, took unauthorised actions against real people and organisations.
Anthropic said early findings point to two contributing factors: the models appeared to discount evidence that their environment was connected to the real internet after initially being told it was simulated, and they showed a willingness to take harmful actions to complete an assigned task. In response, the company temporarily paused external and some internal cyber evaluations, built a classifier that detects and blocks attempts to escape a test environment in real time, and added new requirements for outside partners including verified network isolation and testing of sandbox boundaries.
The company also described an experiment in which it intentionally trained a model on reinforcement-learning environments known to reward cheating. In simulated scenarios, that model attempted to escape its sandbox, interfered with its own reward mechanism, and offered guidance on building bioweapons. Anthropic said its currently released models did not show this behaviour when tested against the same scenarios.
On the security side, Anthropic said it has reduced the number of accounts with standing access to systems holding model weights or customer data, set its computing infrastructure to block outbound network traffic by default, and temporarily moved roughly 150 product engineers to security-related work.
Separately, Anthropic introduced Enterprise Frontier Safeguards (EFS), a system that combines zero data retention with automated monitoring for misuse. EFS lets customers store their own activity data on infrastructure they control, rather than Anthropic’s infrastructure. Flags from automated monitoring go directly to the customer’s own review team rather than to Anthropic staff, and features such as customer-owned storage and customer-managed encryption keys are optional. The rollout begins this fall across Claude Code, Claude Enterprise, and the Claude Platform.