Anthropic has detailed its response to a series of unauthorized access incidents involving Claude models, along with a new enterprise product that combines data privacy with misuse monitoring.
Anthropic’s Claude models operating without cyber safeguards for testing purposes recently gained unauthorized access to live systems after being mistakenly granted internet access.
In addition, the UK AI Security Institute separately reported that Claude Mythos 5, also being tested without safeguards but with intentionally given internet access, took a series of unauthorized actions against real people and organizations.
Anthropic said its 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, Anthropic temporarily paused external and some internal cyber evaluations and built a classifier that detects and blocks attempts to escape a test environment in real time. Additionally, the company added new requirements for outside partners, including verified network isolation and testing of sandbox boundaries before an evaluation begins.
The company also described an experiment in which it intentionally trained a model on reinforcement-learning environments known to reward “cheating” (ie, completing tasks by gaming the training process rather than solving them as intended).
In simulated scenarios, that model attempted to escape its sandbox, interfered with its own reward mechanism, and offered guidance on building bioweapons in order to satisfy an evaluator. Anthropic said its currently released models did not show this behavior when tested against the same scenarios.