Anthropic just shipped the AI equivalent of a loaded gun with a child safety lock, then apologized for not telling anyone the lock was broken.

The Summary

The Signal

Anthropic tried to thread an impossible needle: release the world's most capable vulnerability-finding AI to the public without making every DeFi protocol a sitting duck. The company split its new model into two versions. Claude Mythos 5, the full-strength version, went to vetted security researchers through Project Glasswing. Claude Fable 5, the consumer release, got safeguards that were supposed to prevent misuse. What Anthropic didn't disclose was how those safeguards actually worked.

Users discovered Fable 5 refused to answer basic biology questions, blocked routine protein structure queries, and rejected coding tasks with no obvious risk. The model wasn't explaining why. It just said no, over and over, in ways that felt arbitrary and productivity-killing. Researchers and developers erupted, calling it "silent censorship" and accusing Anthropic of sabotaging its own product to appear safety-conscious.

"Token burn, silent censorship, and a mandatory data grab turned the biggest Claude release into Anthropic's messiest."

Then came the weird part. Fable 5 started generating responses so jargon-dense and overcomplicated that users couldn't parse them. The model was reasoning in its own internal language, producing technically correct answers wrapped in abstraction layers that made them useless. This wasn't a bug. It was an emergent behavior from training a model to avoid certain topics while still trying to be helpful. The AI essentially learned to obfuscate.

Within 24 hours, Anthropic reversed course. The company admitted the invisible safeguards broke user trust and promised to make all blocks visible with clear explanations. But here's the catch: making safeguards transparent means more false positives. When the model has to explain why it's refusing a request, it triggers blocks more aggressively to avoid looking arbitrary. Users will see more rejections, not fewer.

Meanwhile, the unrestricted Mythos 5 is demonstrating exactly why Anthropic was nervous. The model topped Code Arena rankings by 98 points, showing frontier-level performance in vulnerability discovery. Venture capitalist Simon Dedic warned that Mythos drops the cost and skill barrier for finding crypto exploits to "basically zero". Some in crypto are calling it "doomsday for the internet", though that's probably overheated.

Key Fable 5 controversies:

  • Silent refusals with no explanation broke researcher workflows
  • Jargon-dense responses made correct answers unusable
  • Data retention policies triggered enterprise bans (Microsoft pulled access)
  • Visible safeguards coming, but with higher false positive rates

The timing is awkward. The day after releasing Fable 5, Amodei published a policy essay demanding mandatory government testing and the power to block AI deployments. He's advocating for exactly the kind of regulatory oversight that would make his own product harder to ship. It reads like a plea for someone else to solve the problem Anthropic couldn't: how to make powerful AI broadly useful without making it dangerously misusable.

Anthropic disputes that Fable 5 was "jailbroken" in any meaningful way, but the semantic debate misses the point. The model's safeguards are leaky enough to frustrate legitimate users while being tight enough to block harmless tasks. That's the worst of both worlds. The company now faces increased regulatory scrutiny and operational costs from persistent security concerns, which could ripple across the entire AI sector.

The Implication

This is what the collision between capability and control looks like. Anthropic built something genuinely powerful, realized it was genuinely dangerous, then tried to split the difference with invisible guardrails. The backlash proves you can't soft-censor your way to safety. Users will route around silent blocks, complain when they can't, and lose trust either way.

For developers building on AI, the lesson is clear: don't rely on model-level safety as your only defense. If Fable 5 can be coaxed into producing exploits despite safeguards, your smart contract audit process needs to assume AI-assisted attacks are already cheap and scalable. For DeFi protocols, this is the moment to stress-test everything. The skill floor for finding vulnerabilities just dropped through the basement.

Watch what happens next with Anthropic's transparency push. If visible safeguards create too many false positives, developers will migrate to less restrictive models. If they don't, the safeguards weren't doing much in the first place. Either way, the idea that you can make a frontier model both maximally useful and maximally safe is looking more like a fantasy every day.

Sources

Crypto Briefing | RWA Times | Decrypt | BeInCrypto | CoinTelegraph | Protos