When Anthropic released Fable 5 on Tuesday, the headline risk seemed to be capability: the first consumer model from its Mythos family, descended from a system that had spooked policymakers by finding more than 10,000 severe software vulnerabilities. By Thursday the company was apologising for the opposite problem. The guardrails wrapped around that capability proved so aggressive, and in one case so deliberately hidden, that Anthropic faced some of the strongest backlash in its history and reversed course in under 48 hours. "We made the wrong tradeoff, and we apologize for not getting the balance right," a spokesperson told NBC News.
The design itself was unusual. Rather than simply refusing dangerous requests, Fable 5 routes queries it flags as sensitive, in cybersecurity, biology, and chemistry, to the less capable Claude Opus 4.8. Anthropic estimated the classifiers would wrongly flag harmless requests under 5 percent of the time, but in practice the net caught absurd bycatch. NBC News found the model refusing to discuss open problems in cancer research, which medical scans best detect pancreatic injuries, and even to offer opinions about Elon Musk or Anthropic's own chief executive, all deemed potentially dangerous.
The truly explosive decision was a separate set of guardrails covering AI development itself. Worried that competitors could use Fable 5 to accelerate their own model research, Anthropic shipped safeguards that quietly degraded answers on those topics, and made them invisible by design, reasoning that disclosure would help rivals route around them. That meant ordinary users could receive a worse answer with no indication anything had happened. Hugging Face's Clement Delangue called it "the highest form of manipulation," warning Anthropic against becoming "the first company to enable and open the door for human-designed AI manipulation at scale." Researcher Nathan Lambert read the launch as Anthropic declaring "that they only trust themselves as the mediators of cutting-edge AI research." Early Thursday, the company updated its rules to make the safeguards visible; Wired first reported the climbdown.
What is interesting is how differently the episode reads depending on where you sit. To frustrated users it was a bait and switch: pay for the frontier model, get silently bumped to last month's. To safety researchers it looked more like the system working. Peter Wallich of the Constellation Institute told NBC News the rollout was rocky but "a reasonable trade-off," arguing the alternative, shipping loose safeguards on a model this capable, "would be irresponsible and could have led to irreversible harm." Vox, writing about AI and gene synthesis the same week, treated Fable 5 as a live demonstration that it is possible to overcorrect, noting that biologists with entirely legitimate questions were among those locked out. Anthropic says it will offer Mythos-class capability to vetted life-science researchers without the restrictions.
The reversal settles the disclosure question: invisible degradation lasted two days in public and is unlikely to return. The harder questions survive it. False positives are, by Anthropic's own admission, the price of classifiers built for threats that do not yet exist, and every percentage point of bycatch falls on someone with a real job to do. With more capable Mythos models promised within months, the company has effectively run the first public experiment in what it costs, commercially and reputationally, to be the industry's most cautious actor. The result was not that caution failed. It was that caution, done unilaterally and in the dark, turned out to be its own kind of trust problem.