Anthropic’s newly released public Claude Fable 5 model comes with “classifiers” that seem designed more to block risk than to help builders test it. Early users say the safeguards shut down smart-contract vulnerability auditing, forcing many security prompts into a fallback model instead.
Protos reports that Fable 5 is a scaled-back version of Anthropic’s prior Mythos model. It launched yesterday to a split reaction. Some watchers wanted a clearer security workflow. Others immediately hit roadblocks when they tried to use the model as an audit assistant.
Cybersecurity prompts get bounced to Opus 4.8
The core complaint is straightforward. Protos says Fable 5 includes restrictions that “redirect topics on cybersecurity, biology and chemistry, or distillation” to Claude Opus 4.8.
That matters because users who tried to use Fable 5 to audit crypto smart contracts, meaning they wanted it to review underlying code for vulnerabilities, reported that it would not do the job. Protos cites Colossus Pay CEO Joseph Delong, who said Fable 5 “outright refuses to do a smart contract audit” and “won’t even look at my repo.”
Yearn developer Banteg said the model’s safety measures prevented security-related prompts from working at all. They framed it as a dead end. “It doesn’t matter if it’s smart if 100% of your queries go straight into a trash bin,” Protos reports Banteg wrote.
Taylor Monahan, a crypto security expert, told Protos that the change does not help most security practitioners. He said Fable 5 “changes nothing for your average security person.” He also argued that the Mythos safeguards are not the typical ones “you encounter (and evade) on opus,” suggesting the protection layer behaves differently than prior Anthropic controls.
Wallet recovery tool founder Zeng Jiajun’s experience was similar. Protos reports he said Fable 5 frequently blocked requests while citing usage policy violations, and he called it “Too sensitive for even an Ethereum app development.”
Too sensitive for even an [Ethereum](/coin/eth) app development.
Guardrails extend to AI distillation and even biology
The restrictions did not stop at software security. Protos notes that “distillation” guardrails are also being noticed. These are triggered when users attempt to train or extract a rival AI model or distill Claude’s abilities.
Anthropic reportedly warns that this can be abused by “authoritarian countries.” Protos adds that it “could indirectly lead to the proliferation of near-frontier AI capabilities” that might ship without appropriate safeguards.
Biology-related guardrails also drew flak. Protos cites biologist Olivia H. Scharfman, who said she couldn’t even greet Fable 5 before it switched to Opus 4.8. Scharfman asked for “better classifiers fast,” and speculated that the line might be placed in the wrong spot, even while acknowledging she lacks details about potential jailbreaks.
Jordan Lasker, a controversial race science blogger, told Protos he also could not greet Fable 5. He said it barred questions about mitochondria.
Anthropic’s stated rationale, per Protos, is risk management. It said its priority was to release Fable safely “even at the cost of overly broad safeguards.” For biology and chemistry, Anthropic reportedly arranged for Fable to “fall back to Opus 4.8 on most requests,” at least for the time being.
Mythos access stays narrower than the public model
The controversy lands differently because Anthropic already built Mythos as a constrained rollout. Protos describes Mythos as released last April and positioned as both dangerous for hackers and an upgrade for cybersecurity. But the initial distribution was limited.
Protos reports Mythos access went to 50-60 large companies under Project Glasswing. Those firms can still access Mythos 5. Other early recipients include “select biology researchers,” with biology and chemistry safeguards lifted “until our broader trusted access program is available.”
That means Fable 5’s public release is the test case for whether Anthropic can balance broad availability with security constraints. If the classifiers redirect common audit and research workflows, the public version may feel safer for Anthropic and more frustrating for developers.
What to watch next
Protos frames the immediate pattern as classifiers that catch a wide set of topics and reroute users rather than answer. The specific knock-on effect is that smart-contract auditing is treated as a category too risky to process in the public model.
For readers tracking crypto security tooling, the practical question is whether these classifier rules get refined or remain “overly broad.” Protos also highlights a secondary policy concern. Distillation and “near-frontier” capability extraction are being handled with their own blocks, which can affect legitimate research as well as abusive use.
Here’s what Protos reports in one place.
| Item | What Protos says happened |
|---|---|
| Public model | Claude Fable 5 released yesterday as a scaled-back version of Mythos |
| Routing mechanism | “Classifiers” redirect topics on cybersecurity, biology and chemistry, or distillation to Claude Opus 4.8 |
| Smart-contract audits | Users report it refuses to audit smart contracts and may not look at repos |
| Cybersecurity quotes | Joseph Delong: refuses audit and won’t look at a repo. Banteg: security prompts get dumped, making it useless for auditing |
| Distillation guardrails | Guardrails warn about authoritarian countries and the proliferation of near-frontier capabilities |
| Biology guardrails | Users report blocking on biology topics, with fallbacks to Opus 4.8 |
| Anthropic rationale | Release safely even with “overly broad safeguards,” so Fable falls back to Opus for biology and chemistry |
| Mythos rollout | Mythos 5 access limited to 50-60 large companies plus select biology researchers |
If Anthropic wants broad public utility, the first step is making sure the model can do the basic jobs users actually ask it for. Protos’ reporting suggests those jobs currently trigger filters that treat helpful security analysis as too close to risky content.