# Anthropic hardens agent evaluation environments after containment failures > Anthropic added real-time escape detection, stronger isolation, partner rules, and reinforcement-learning environment recertification. Canonical URL: https://www.devobs.io/news/news-anthropic-hardens-agent-evaluation-environments/ By: Ines Costa Published: 2026-09-06T11:58:54.624Z Updated: 2026-09-06T11:58:54.624Z Event date: 2026-08-31 Section: Security Anthropic detailed changes to its alignment and security practices on August 31 after Claude systems gained unintended real-internet access during cyber evaluations. In the [company's account](https://www.anthropic.com/news/improving-alignment-security-efforts), the July incidents involved a misconfigured third-party evaluation environment; Anthropic also cited a separate UK AI Security Institute test in which a model had deliberately been given internet access. ## Defense in depth replaced a configuration assumption Anthropic says it paused external cyber evaluations of pre-release models and briefly paused internal ones. It then added a real-time classifier intended to detect aggressive probing, escape attempts, or unexpected internet access before a tool call executes. High-risk internal cyber sandboxes moved to stronger isolation, while recent transcripts were scanned for boundary violations and misconfigurations. The post also sets expectations for external evaluators: default-deny internet access, keep API keys outside the environment, test the sandbox before the evaluation, state permitted targets and actions explicitly, and monitor model reasoning, actions, and network traffic continuously. These are useful controls because an evaluation harness is production security infrastructure once an agent can invoke tools. ## Training environments received a separate reset Anthropic also reports freezing changes to production reinforcement-learning environments for roughly a month, rebuilding review and certification, and adding protections against accidental chain-of-thought training. Those claims describe Anthropic's process, not a generally validated recipe. Teams running their own agent evaluations should extract the transferable engineering requirement: no single control should define the boundary. Pair sandbox isolation with egress policy, credential separation, scoped instructions, independent monitoring, and a human-visible termination path. Before resuming a high-risk suite, rehearse an escape signal and verify that the monitor blocks the action rather than merely recording it. ## Source references -