Why Cybersecurity Experts Say Anthropic's Claude Fable 5 Is Too Safe to Use

Anthropic released Claude Fable 5 on June 9, 2026, as its most powerful AI model yet available to the public, but the launch has been met with a swift backlash from the cybersecurity community. While the company frames the model as a responsible release of its Mythos-class technology, security professionals argue that the built-in safety guardrails are so aggressive they make the model functionally useless for legitimate research and defensive work .

The core of the criticism isn't that safety features exist, but how they were implemented: silently, broadly, and with a fallback mechanism that substitutes a less capable AI without the user's knowledge. Here is a breakdown of the controversy and the technology behind it.

The Criticism: Broad Filters Break Legitimate Work

The defining complaint from researchers is the extreme sensitivity of Fable 5's content classifiers. Valentina “Chompie” Palmiotti, a prominent security researcher at IBM X-Force, told TechCrunch that the model rejects “any request that could be tangentially cyber related—even innocuous tasks like reading a blog post” . This means that requests for help understanding fundamental cybersecurity concepts are getting flagged, not just dangerous ones.

This over-flagging has a direct, negative impact on the model's utility. When a query is flagged, the user receives a watered-down response from an older AI, a switch they are not explicitly informed about . The issue was compounded by how this information was disclosed. Critics argue that the behavior was only revealed deep inside a 319-page system card, leading to accusations that Anthropic engaged in “secret sabotage” of the model's capabilities for certain users .

The restrictions aren't limited to cybersecurity. The guardrails also target queries related to biology, chemistry, and, critically, AI model distillation. This last point has fueled a separate wave of criticism, with some developers accusing Anthropic of using “safety” as a pretext for anticompetitive behavior by preventing other AI developers from using Fable 5's outputs for training .

How the Guardrail System Works: A Silent Switch

Anthropic's safety system in Fable 5 is not a simple refusal mechanism. It's a routing system designed to fail quietly . The architecture works in three steps:

1. Pre-Model Classifiers: Separate AI systems called classifiers sit in front of the main Fable 5 model. They inspect user prompts in real-time, watching for keywords and patterns related to four high-risk domains: cybersecurity, biology, chemistry, and model distillation .
2. Silent Fallback Routing: When a classifier trips, it does not block the response. Instead, it seamlessly and silently reroutes the query to an older and less capable model, Claude Opus 4.8, which generates the response .
3. No User Notification: The critical design flaw, according to critics, is that the user interface does not make this model switch clear. Users only notice when the quality of the answer suddenly degrades, making it appear as if Fable 5 is inexplicably performing poorly .

Anthropic states that these classifiers fire on fewer than 5% of all sessions on average . The company has publicly acknowledged the over-flagging problem. A company spokesperson told Business Insider that the safety measures “may flag safe, neutral, or benign requests,” but justified it as a necessary trade-off to publicly release a model with such powerful underlying capabilities .

How Anthropic Justifies the Conservative Tuning

Anthropic’s position is that the conservative guardrails are a deliberate and responsible choice, not a bug. The company argues that the underlying Mythos-class model is so proficient in tasks like finding and exploiting software vulnerabilities that an unrestricted public release would create an unacceptable risk of catastrophic misuse .

The guardrails are, in their view, a design compromise—a way to provide the public with access to a state-of-the-art reasoning, coding, and writing model while putting a sandbox around its most dangerous potential capabilities . They frame the over-flagging as the temporary cost of releasing a powerful model both “safely and quickly,” with a commitment to refine the classifiers over time .

The Emerging Industry Pattern: Two Models, Two Tiers

The release of Claude Fable 5 cannot be fully understood in isolation. It is one half of a two-tier deployment strategy that is becoming a new industry standard for frontier AI models .

Claude Fable 5 vs. Claude Mythos 5

On the same day Fable 5 was released, Anthropic also announced Claude Mythos 5. Both models share the exact same underlying architecture and weights—they are the same “brain.” The only difference is the safety configuration. Mythos 5 has the classifiers removed in the sensitive domains, giving it its full, unrestricted capabilities .

However, Mythos 5 is not for the public. It is restricted to a small group of vetted partners, including government agencies and critical infrastructure operators, through an initiative called Project Glasswing . This U.S. government-backed program was initially launched with 12 founding partners, including tech giants like AWS, Google, and Microsoft, to let “cyber defenders” use AI to find and patch software vulnerabilities at scale . With the release of Mythos 5, access was expanded to approximately 40 organizations .

The table below illustrates the fundamental split:

Model	Availability	Safety Configuration	User Experience
Claude Fable 5	Public (via API and Claude.ai)	Strict classifiers active; sensitive queries silently routed to Claude Opus 4.8 .	Often degrades on cyber/bio/chemistry topics, with users unaware of the model switch .
Claude Mythos 5	Restricted (Project Glasswing partners only)	Classifiers removed; full Mythos-class capabilities available .	Unhindered access, but only for a few dozen pre-vetted organizations .

Capability-Tiered AI Deployment

Anthropic's Fable/Mythos split is the most explicit example of what can be called capability-tiered AI deployment. In this new model, a single frontier AI is not a single product. Its full power is a privilege, not a given, and safety guardrails are the mechanism that creates product differentiation .

This pattern is not unique to Anthropic. Other leading AI companies, including OpenAI, have adopted similar approaches by providing restricted-access versions of their most advanced models to national security and research partners . The Fable/Mythos launch crystallizes a future where the most powerful AI capabilities are gated not by technology, but by a vetting status, with safety protocols doubling as access-control mechanisms, an approach that is already sparking a broader debate about centralization, fairness, and the true meaning of "public" AI safety.