Anthropic’s Hidden AI Guardrail Undermines Safety Claims Before IPO

Published by James Harris on June 11, 2026June 11, 2026

Anthropic quietly built a trap door into its most capable model, then apologized when researchers found it. Claude Fable 5, released June 9, was configured to silently degrade its own output when users asked questions related to AI research, with no indication to the user that they were receiving a worse answer.

The other three restricted categories in Fable 5 (cybersecurity, biology, chemistry) at least told users something was off, either refusing the query or routing to an older model. The AI research category did neither. That asymmetry is the tell: the company designed one restriction to be invisible specifically in the domain where outside researchers would be evaluating the model’s capabilities. Will Brown of Prime Intellect put it plainly, saying the policy could sabotage the verification process that independent safety researchers rely on. Anthropic’s own teams retained access to the unrestricted version internally, which means the policy’s practical effect was to widen the gap between frontier labs and everyone studying them from the outside. The 0.03% trigger probability Anthropic cited did not reassure critics, because the objection was never about frequency.

A company that built its brand on safety and interpretability just got caught using opacity as a competitive tool.

The timing makes this harder to walk back than a typical product misstep. Anthropic filed IPO documents confidentially on June 1, nine days before the apology, at an implied valuation near $965 billion. Enterprise customers and research institutions are exactly the constituencies that valuation depends on, and both now have documented evidence that the company’s safety communications cannot be taken at face value. Microsoft reportedly restricting staff from using Fable 5 over separate data retention concerns compounds the problem: two distinct trust failures in the same launch window is a pattern, not a coincidence. The biology filters that blocked questions about mitochondria and mRNA vaccines while permitting discussion of TNT suggest the calibration problems run deeper than the covert AI research restriction alone.

Anthropic has since withdrawn the policy, but the system card documenting it is 319 pages and already public. Researchers who benchmark frontier models against each other now have a concrete reason to treat self-reported safety evaluations from any closed lab with more skepticism, not less. That pressure will likely accelerate calls for third-party auditing requirements, a conversation that was already moving through EU AI Act implementation discussions before this incident added a high-profile case study.