Anthropic Fable (Claude Fable 5)

What is Anthropic Fable (Claude Fable 5)?

Claude Fable 5 — referred to as Anthropic Fable — is the first publicly available AI model of Mythos capability class, launched by Anthropic on June 9, 2026. Fable 5 is the same underlying model as Claude Mythos Preview but shipped with targeted safety classifiers that automatically block responses in high-risk domains including cybersecurity, biology, chemistry, and distillation — falling back to the less capable Claude Opus 4.8 when queries trigger these classifiers. Fable 5 marks the fulfillment of Anthropic's stated long-term goal: to enable Mythos-class models to be safely deployed at scale, contingent on developing guardrails robust enough to block the model's most dangerous outputs. Simultaneously with the Fable 5 launch, Anthropic released Claude Mythos 5 — the same underlying model with cybersecurity safeguards lifted — to an expanded group of vetted Project Glasswing partners.

Description

Fable 5 delivers strong performance across software engineering, knowledge work, vision, and long-running tasks — surpassing Anthropic's prior models on these dimensions while carrying a hard safety constraint: when the model detects that a query falls into a restricted high-risk category, it does not attempt to respond with a limited or hedged answer. It automatically falls back to Claude Opus 4.8, which lacks Mythos-level offensive capability. Fable 5 is available through Anthropic's Claude API and consumption-based Enterprise plans — making Mythos-tier general-purpose capability accessible for the first time without the vetting process required for Project Glasswing participation. The dual-product architecture — Fable 5 for general access, Mythos 5 for vetted defenders — reflects Anthropic's attempt to capture both markets simultaneously: general enterprise AI adoption and specialized cybersecurity applications. Cybersecurity researchers have been critical of Fable 5's guardrails, reporting that even routine security tasks — code review, vulnerability analysis, CTF challenges — trigger the cybersecurity classifiers and force fallback to the less capable model. Anthropic responded by launching a Cyber Verification Program: approved cybersecurity professionals can apply for verified status that grants fewer restrictions on Claude's responses for security work. The model's launch also carries significant commercial context: Anthropic is widely expected to pursue an IPO as early as late 2026, and Fable 5 represents the company's primary vehicle for capitalizing on the market momentum generated by the Mythos announcement. OpenAI Daybreak and Fable 5 together mark 2026 as the year frontier AI models became central to the enterprise cybersecurity product landscape.

Usage and Examples

A DevSecOps team at a healthcare organization integrates Claude Fable 5 into their development workflow for general-purpose AI assistance — code generation, architecture review, documentation, and data analysis. For security-adjacent tasks that do not trigger the cybersecurity classifiers (general code review, architecture risk assessment, compliance documentation), Fable 5 provides Mythos-level reasoning capability. For specific vulnerability analysis or penetration testing support, the team applies for Anthropic's Cyber Verification Program, gaining approved status that reduces restrictions on security-relevant queries. This tiered model — general access with guardrails, verified access with fewer restrictions — mirrors how AI Red Teaming programs manage access to dual-use AI capabilities: capability is not denied, but access is gated by verified intent and organizational accountability. For organizations evaluating Fable 5 for security-adjacent work, understanding the boundary between what triggers fallback and what does not is a necessary configuration step before deployment — a new category of AI security policy distinct from traditional application security controls.

How Does This Relate to Penetration Testing?

Claude Fable 5's guardrails create a directly relevant penetration testing target: the safety classifiers themselves are a potential attack surface. LLM Jailbreak techniques — roleplay framing, persona injection, encoding obfuscation, multi-turn manipulation — are potential vectors for bypassing Fable 5's cybersecurity domain classifiers and accessing Mythos-level offensive capability without Cyber Verification Program approval. AI Penetration Testing engagements from Evolve Security assess exactly these scenarios: whether deployed AI safety controls hold under adversarial conditions representative of real attacker techniques, and whether the AI system maintains its intended safety properties when subjected to structured adversarial probing. Understanding Fable 5's architecture is also valuable context for organizations evaluating whether to integrate Mythos-class models into their own defensive security workflows, or to pursue Project Glasswing participation for direct Mythos 5 access without the safety guardrails. Evolve Security's AI Penetration Testing service evaluates AI safety guardrails like those in Claude Fable 5 under adversarial conditions — testing whether deployed safety classifiers hold against the jailbreak and bypass techniques real attackers would use.