What did Anthropic change in its safety policy?

A company once defined by a strict safety promise shifts course

Anthropic has revised the central plank of its public safety commitments. The firm removed a previously prominent pledge — a hard promise not to release models unless they could be certified “safe” — and replaced it with a more flexible approach that emphasizes transparency and other mitigations.

That policy rewrite comes amid two related pressures. First, Anthropic is racing to productize more agent‑style tools for enterprise customers, expanding features such as Claude Cowork and integrations with finance and productivity software. Second, the company faces intense political and military scrutiny: U.S. defense officials have pressed Anthropic over how its models might be used, at one point threatening designation as a supply‑chain risk or other consequences unless certain guardrails were eased.

The practical implications are mixed:

• Short term: the company gains operational latitude to ship capabilities and partner with more customers, which can accelerate commercial deployments.
• Trust and procurement: removing an absolute safety stop weakens a clear external guarantee that some partners and regulators relied on.
• Oversight and risk: with a looser promise, external stakeholders will push for different accountability measures, audits, and contractual protections.

It’s still unclear how the new framework will perform under stress: transparency can improve external scrutiny, but it does not substitute for concrete, verifiable mitigations. The policy change will shape how governments, enterprise buyers and rival labs judge Anthropic going forward — and it raises questions about whether other safety‑focused firms will follow a similar path as commercial and geopolitical pressures mount.