Anthropic accused of ‘secret sabotage’ as Claude Fable 5 silently limits capabilities for AI researchers and developers

A paragraph buried in Fable 5’s 319-page system card revealed the model would silently downgrade its responses for certain AI development work—without telling users.

When Anthropic made its first Mythos-tier model available to the general public yesterday, called Claude Fable 5, Fortune reported it was a “considerable step” for the lab, coming just over a week after the company confidentially filed for IPO paperwork. It had initially deemed Mythos-class models too dangerous to release, citing their significantly enhanced ability to identify software vulnerabilities, but said it was now confident new guardrails in Claude Fable 5 are enough to ensure these dangerous skills don’t fall into the wrong hands.

Just hours after the model’s release, however, major backlash from AI researchers, developers, and policy experts began brewing on social media. The pushback centered around a paragraph buried in Claude Fable 5’s 319-page system card—a document that offers detailed safety disclosures—which revealed that Fable would quietly downgrade its own responses when it detected requests related to cutting-edge AI development work, such as building the infrastructure used to train large AI models.

In practice, that means a user could ask Fable for help, receive a deliberately weakened answer, but not know the model was holding anything back. Critics made it clear they felt this undermined a basic expectation that a tool would either do what it was asked or tell the user it wouldn’t.

Unlike Fable’s other restrictions, such as around cybersecurity and biology, which openly redirect users to a less powerful model with a visible notification, the system card emphasized that this is “not visible to the user.” The model still responds, but uses “interventions to limit Claude’s effectiveness” without telling the user it’s doing so.

Anthropic estimated the restrictions would affect roughly 0.03% of traffic. But it also defended its effort by saying “enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms.” 

Pushback from AI community

A wide swath of the AI community pushed back sharply—including open-source researchers critical of Anthropic’s closed policies, as well as AI safety experts who typically align with Anthropic.

“To have my access to the cutting edge models for my work rug pulled in an under the table fashion is appalling,” wrote Nathan Lambert, an open-model researcher who most recently led work at AI2. “To me this paints Anthropic clearly as anti-science, and therefore anti-progress and anti-safety.” 

Dean Ball, a senior fellow at the Foundation for American Innovation who previously served as senior policy advisor at the White House Office of Science and Technology Policy, wrote that Anthropic’s “secret sabotage” safety policy “massively and profoundly raises the status of the argument that AI safety has been hype to justify monopolistic behavior by labs.” 

And Jeremy Howard, head of nonprofit research group Fast AI, wrote that “Anthropic has chosen the opposite of the safe path: they are allowing themselves, the current top lab, to use their top model for frontier AI research. They’ve said they’ll sabotage others who try. This means the AI frontier advances, & power imbalance increases.” 

Even former Anthropic employees joined in. Behnam Neyshabur, who previously co-led Anthropic’s effort to develop an AI scientist, posted on X saying: “Working on AI for cancer? Sorry, I can’t help you. Working on AI for Alzheimer’s Disease? Sorry, I’m becoming a bit dumb when it comes to the AI part of it.” In another post, he added: “I’ve argued for the last eight months that this was the direction things were heading. In my view, concentrating these capabilities fundamentally slows scientific and technological progress and is net negative for humanity.”

Not all prominent AI voices weighed in with criticism, however. Ethan Mollick, an associate professor at Wharton studying AI, innovation, and entrepreneurship, did not focus on the restrictions, writing in a blog post that Claude Fable 5 “outperformed basically every other public model I have used by a considerable margin.” 

Former OpenAI cofounder and Tesla AI director Andrej Karpathy, who announced he had joined Anthropic last month, called Claude Fable 5 a “super exciting release” on X and said it is a “major-version-bump-deserving step change forward.” He did, however, point out that the model “still has quirks that people will run into and the safeguards are configured to be a little too trigger-happy for launch, which can hopefully be tuned over time.”  

Anthropic says it wants to make models accessible and safe

Before the release, Anthropic seemed to gird itself for backlash, though it did not specifically address potential blowback regarding the research restrictions. In an interview with Fortune yesterday, Dianne Na Penn, Anthropic’s head of product management, research, and labs, said that the new model was able to produce frontier performance that was 10 to 20 points more than its previous model, Opus 4.8 or other frontier models.

“I think generally being able to do that, at the same time having the right guardrails in place to make it accessible, and generally in a safe manner, I think that’s probably the main thing that I want folks to take away,” she said. “We’re raising the bar on the intelligence of the models, and at the same time, we are pushing the frontier in a safe manner.” 

She added that Anthropic recognized that some benign requests would initially be blocked. “We’re working actively on making those safeguard improvements post-launch, but we wanted to make the model accessible generally in a safe manner as soon as we could.”

Anthropic did not respond to Fortune’s request for comment.

This story was originally featured on Fortune.com