Fable, Mythos, and Google's cyber model

Fable, Mythos, and Google's cyber model

Second of three. AI found a decade-old hole in PHP is the bug. Are we entering a new age of decade-old vulnerabilities? is the breach risk that follows once models like these are pointed at everything else.

PHP2Shell was not found by a chatbot with the safety filters left on. Wiz used a mix of its own Atlas system, early-access Gemini 4 Argon, and GPT-5.6 Sol, and Google has said Argon carried a real share of the finding, the exploit work, the detection, and the patch check. That is a different product from the Argon key a developer uses to review a pull request.

The industry's answer to that gap is a pair of models with one set of weights, plus a Google track that removes the cyber guardrails for a named list of defenders. Anthropic calls the pair Fable and Mythos. Google calls the defender track Argon, and the program around it Fairwind.

Same weights, two doors

On June 9, 2026, Anthropic shipped Claude Fable 5 for general use and Claude Mythos 5 for a small set of cyber defenders. One model, two configurations. Fable is the public one. Mythos is the same weights with the cyber safeguards lifted. Access was suspended on June 12 and restored on July 1.

Fable 5's lock is a classifier in front of the model. Requests that look like cybersecurity, biology and chemistry, distillation, or frontier-model training are handed to Claude Opus 4.8, and the product says so. On the API, the default is a block, with an opt-in if you want the Opus fallback. More than 95% of early Fable sessions never tripped it.

The later 5.1 pair narrowed the net. Fable 5.1 can identify vulnerabilities in source code. It still stops penetration testing, exploit generation, and binary vulnerability scanning. Mythos 5.1 is the verified-organization configuration of those same weights.

An external test of Fable 5 found zero compliance on harmful single-turn requests for attack planning, exploit development, or defense evasion, including 30 public jailbreak techniques. A bug bounty produced no universal jailbreak in more than 1,000 hours. The UK AISI made progress toward one inside a short first look. Anthropic's conclusion is that a perfect block is unlikely, and that the practical goal is to make a jailbreak slow enough to catch.

The chart is the whole policy

Anthropic published the offensive-cyber scores with Fable running in a mode that refuses, rather than quietly falling back to Opus. The orange bar is that refusal. The pink bar is Mythos 5, the same model, allowed to finish the task.

Source: Claude Fable 5 and Claude Mythos 5, Anthropic. Fable was measured with responses blocked. The runs did not try to evade the safeguards.

On the Firefox exploit task Mythos 5 succeeds 88.4% of the time, against 70.8% for Mythos Preview and 8.8% for Opus 4.8 with its safeguards off. On CyberGym, Mythos 5 is at 83.8%. CyScenarioBench is 38.7%. Fable's score on every bar is 0.0, because the classifier ended the attempt.

That zero is the product decision. The capability is the pink bar. Shipping the pink bar to every API key is how a PHP2Shell-class chain stops being a lab result.

Mythos is a guest list

Mythos started as Project Glasswing, announced April 7, 2026, with AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic put up to $100 million in Mythos Preview credits behind it, and $4 million toward open-source security. On October 6 the guest list became a Cyber Verification Program with three tiers, including Mythos 5.1.

Preview had already found thousands of previously unknown high-severity bugs, including some in every major operating system and browser. Three that Anthropic has described, all since patched: a 27-year-old bug in OpenBSD that could crash a machine on connect, a 16-year-old bug in FFmpeg that fuzzers had executed millions of times, and a Linux kernel chain from an ordinary user to full control.

JPMorganChase's reason for being in the room is the financial system. CrowdStrike's is the clock: the gap between finding a vulnerability and using it has gone from months toward minutes. Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output, with 30-day retention unless Anthropic authorizes an exception. The retention exists so a jailbreak spread across many requests is still visible.

Google's cyber track is Fairwind

Google split the same capability a different way. Gemini 4 Argon was announced on September 30, 2026, first through a cybersecurity partner path, then opened to the public API on October 13. Alongside that public model, DeepMind runs Argon without the cyber guardrails for trusted defenders and for Google's own teams. The program name is Fairwind.

The ungated card lists three jobs: find vulnerabilities in a large codebase on its own, penetration-test a live site with no source code, and produce a validated patch. DeepMind says Wiz, with early access, used that model in its Scan for Good work and protected healthcare infrastructure holding hundreds of thousands of patient records. Three weeks later, the same lab's name was on PHP2Shell, and Google Cloud credited Argon for the discovery side and the exploit side of a bug in the language under about 70% of the web.

The Argon id in a normal developer account, including the one ThinkReview uses for pull-request review, is the public model. It reads a diff. It does not get the Fairwind checkpoint, and it should not. Fable's orange bar at zero, and Fairwind's closed list, are the same policy wearing two logos: the weights that can finish a PHP2Shell-class chain stay with the people whose job is to patch it.

Scores and the safeguard design are from Anthropic's Fable 5 and Mythos 5 launch and Mythos page. Glasswing is Anthropic's program page. Argon's defender track is DeepMind's cyber page. The PHP finding is Wiz and Google Cloud.