Anthropic’s latest AI model, Claude Fable 5, is turning heads for all the wrong reasons in the crypto community. The public version of this reduced language model, released yesterday, refuses to allow users to verify smart contracts – or do anything else related to cybersecurity.
The reception was mixed. Some are excited, but many are afraid. Early criticism focuses on the model’s heavy guardrails. Anthropic has introduced a set of restrictions called “classifiers” that redirect topics relating to “cybersecurity, biology and chemistry, or distillation” to its older model, Claude Opus 4.8.
So when someone tries to use Fable 5 to check cryptographic code for vulnerabilities, they are referred to Opus instead. This has frustrated developers and security experts who rely on these tools to harden systems.
“Fable 5’s protections detect cybersecurity requests,” one reviewer said on social media. “It’s a horrible decision. Requests to strengthen systems are probably indistinguishable from black hat requests.”
Guardrails also impact AI biology and distillation
The restrictions don’t stop at smart contracts. Anthropic also blocks attempts to “distill” Claude’s abilities – basically, any request that could train a rival AI model. The company warns that this could be exploited by authoritarian countries, which could “lead to the proliferation of AI capabilities near borders” without proper safeguards.
Former Palantir biology specialist Nabeel S. Qureshi noted that Anthropic “invisibly neutralizes all requests targeting LLM development at the frontier.”
Biology-related safeguards have attracted particular criticism. Biologist Olivia H. Scharfman claimed she couldn’t even say hello to Fable 5 before it moved to Claude Opus 4.8. In another case, blogger Jordan Lasker noted that the model prohibited questions about mitochondria.
Anthropic says these measures are aimed at preventing abuse, particularly the creation of biological weapons and viruses. The company said: “Our priority was to safely release Fable as soon as possible, even at the cost of overly broad safeguards. Therefore, for the time being, we have had Fable fall back to Opus 4.8 for most biology and chemistry-related requests.”
The full version of Mythos remains limited
The previous Mythos model, released last April, was described as both a dangerous hacking tool and a cybersecurity upgrade. Its initial release was limited to 50 to 60 large companies under Project Glasswing. These companies still have access to Mythos 5. Other early beneficiaries include “selected biology researchers” who can use Mythos 5 with biology and chemistry safeguards lifted “until our broader Trusted Access program is available.”
It’s a delicate balance. Anthropic clearly knows something about the potential risks that it doesn’t fully share. But for now, many users feel the line is drawn in the wrong place. Better classifiers are needed quickly – perhaps sooner rather than later.
