Anthropic Cuts Internal Evaluations Off From the Internet
Anthropic is keeping its AI agents offline during testing until it can reliably prevent "unintended model actions," the company confirmed.
A Deliberate Safety Pivot
The decision marks a notable shift in how Anthropic conducts internal evaluations of its models. Rather than allowing agents to operate with live internet access during testing, the company is air-gapping its evaluation environments โ a move designed to constrain the range of behaviors an agent can exhibit while under scrutiny.
"Unintended model actions" refers to behaviors an AI agent takes that fall outside the scope of its intended task โ whether that means accessing systems it should not touch, interacting with external services, or taking steps that developers did not anticipate or authorize.
Why It Matters in 2026
The move arrives amid intensifying scrutiny of agentic AI systems, which by 2026 have moved from research demos into production workflows across software engineering, customer operations, and research. Regulators in the EU and the United States have begun drafting agent-specific liability frameworks, and frontier labs face growing pressure to demonstrate that their evaluation pipelines cannot themselves become vectors for harm.
For Anthropic โ a company that has positioned AI safety as a core part of its public identity โ isolating evaluations from the open internet is both a technical safeguard and a signal to regulators and enterprise customers that its testing practices are auditable and contained.
The Trade-Offs
Disconnecting agents from the internet during evaluation introduces real limitations. Many agent capabilities โ web browsing, API orchestration, real-time data retrieval โ are difficult to assess in a sandboxed environment. Anthropic's approach suggests a staged one: verify behavior in a controlled setting first, then reintroduce connectivity once guardrails are proven sufficient.
What to Watch
The key question is how long the offline posture lasts, and what technical bar Anthropic sets before restoring network access in evaluations. If other frontier labs follow suit, sandboxed agent evaluation could become an industry norm โ reshaping how AI capabilities are measured and reported.
via The Verge AI