Anthropic has reversed a policy it put in place without disclosing it, after researchers said it was actively sabotaging their work.
The policy was not made public. Researchers pushed back hard enough that Anthropic walked it back. What the policy covered, which research it affected, and what specifically it blocked are unclear. The word researchers reached for was "sabotage," which suggests the impact was more than inconvenient.
That framing matters because Anthropic invests heavily in positioning itself as the AI lab that takes external research and safety collaboration seriously. A concealed policy that researchers describe as undermining their work is a direct contradiction of that positioning. A company can reverse a bad decision quickly and still have made the bad decision.
The reversal itself is the tell: Anthropic wouldn't walk something back unless the pressure was real. The more interesting question is whether this was a one-off or evidence of a broader pattern of undisclosed policies quietly shaping who can do what with its models.