Skip to main content
November 15, 2026

Claude Fable 5.2 and Mythos 5.2: Safety Tiers for Security Agents

Fable 5.2 vs Mythos 5.2: Cyber Rules for Security Agents

On November 9, 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. The launch page says plainly that "Claude Fable and Claude Mythos are the same underlying model." The difference is the safeguards. Fable 5.2 is the general release, with classifiers that block high-risk cyber and biology work. Mythos 5.2 is the version for vetted organizations in those dual-use fields.

For a team whose agent triages CVEs, reads exploit traffic in access logs or analyzes a binary found in /tmp, this matters: the model you can call by default refuses much of that work. This post maps security tasks to tiers and lists what breaks in an unattended agent when the tier is wrong.

What Anthropic Shipped

Claude Fable 5.2 Claude Mythos 5.2
Access All platforms, API ID claude-fable-5-2 "Limited availability": organizations verified through Anthropic's verification programs
Price per 1M tokens $8 input, $40 output $8 input, $40 output
Fast mode $16 input, $80 output Only for organizations in the Verification Programs
Cyber safeguards Classifiers block high-risk cyber activity Reduced, by verification tier
Context and output 1M context, 128K max output 1M context, 128K max output

The standard prices, context and output limits come from the models overview; the fast mode rows come from the launch page. The launch page adds that both models "generate outputs up to 30% faster" than the 5.1 versions, and that fast mode "speeds up token generation by up to 2.5 times" at twice the standard price.

What "Hardened" Means in Practice

The launch page lists what the Fable 5.2 classifiers block: high-risk cyber activities "including penetration testing, exploit generation" and binary-based vulnerability work. It is also direct about the scope: "Claude Fable 5.2's safeguards will block most cybersecurity-related tasks."

In early testing across claude.ai, Claude Code and Cowork, Anthropic measured about 60% fewer cybersecurity blocks than with Fable 5.1. Fewer is not zero, and the page counts blocks in Anthropic's own apps, not in your agent's prompts.

When a query is blocked, the launch page says users get a response from Opus 4.8 for cybersecurity instead. On the API, this handoff is not automatic. The refusals and fallback docs describe server-side fallback as a beta that applies only when the request sets the fallbacks parameter. Without it, a classifier block comes back as HTTP 200 with stop_reason: "refusal", and stop_details.category can be cyber. With it, the top-level model field names the model that actually answered. A cyber refusal before any output is not billed, but a refusal in the middle of a stream bills the input and the output already streamed.

The Cyber Verification Program

Mythos 5.2 access goes through the Cyber Verification Program (CVP). The support page says it is for verified security professionals who want the most capable models "with fewer interruptions from our cyber safeguards." It has three tiers:

Tier Covers Who can apply Target review time
Defense Access Security operations, incident response, malware analysis, detection engineering, vulnerability triage and validation Organizations and individuals Two days
Red Team Access Authorized penetration testing, red teaming, adversary emulation, offensive tool development, exploit development Organizations only Two weeks, currently longer because of demand
Specialized Access Authorized pen testing and red teaming of critical infrastructure, and research into the highest-severity techniques Organizations only No set target

Each tier includes Mythos, Opus and Sonnet models, starting with Mythos 5.2, Mythos 5.1, Opus 5.5 and Sonnet 5.5, with safeguards adjusted by tier. The page does not describe reduced safeguards for Fable. So do not expect verification to make your Fable 5.2 agent less strict. Plan to get the verified behavior by switching the agent to a covered model.

Some things stay prohibited in every tier, including denial-of-service attacks, mass destruction of data or devices, extortion and "any request with stated malicious intent." The Usage Policy still applies in full.

Which Tier Fits Which Agent Task

The CVP tier descriptions and the Fable block list give a rough map. Treat it as a starting point, and test your real prompts.

Agent task Fable 5.2 without verification Tier to plan for
Disk-full, OOM or latency alert triage Not a cyber task None
Summarize a CVE advisory and match it to your package inventory Can be flagged as cyber Defense Access
Read a WAF or access log full of exploit attempts during an incident Can be flagged as cyber Defense Access
Analyze a suspicious binary or decoded payload from a host Binary-based vulnerability work is on the block list Defense Access (malware analysis)
Confirm that a CVE is exploitable in your staging environment Exploit generation is on the block list Red Team Access, organization only
Run an authorized penetration test on your own services Penetration testing is on the block list Red Team Access
Scan your own repositories for vulnerabilities Mostly blocked CVP, or Claude Security on Claude Enterprise

The last row has a shortcut. The launch page says Claude Enterprise customers can use Claude Security without CVP membership, and that it is powered by the full model family, including Mythos 5.2.

Five Traps for Unattended Agents

The CVP page has rules that look like paperwork but break automation.

1. The Default workspace does not count. On the Claude Console, an admin assigns the program per workspace, and the Default workspace cannot be used. If your agent's API key lives in the Default workspace, approval changes nothing for it. Create a dedicated workspace and key for the security agent.

2. Key rotation is part of the deal. Defense Access requires you to attest to MFA and to API keys replaced at least every seven days. The page also says that starting December 15, 2026, Defense Access requires phishing-resistant MFA and no longer allows static or long-lived API keys. A key pasted into an agent's config file once and never changed does not meet either rule. Check now which credential your agent will need after that date.

3. Zero data retention conflicts with enrollment. Data retention is required for everyone enrolled. ZDR is possible only for organizations with a data retention exemption for Fable or Mythos, and then only for internal use. If your security policy demands ZDR for incident data, settle that before you apply. On Amazon Bedrock, the page says CVP is for now available only to organizations with a data retention exemption, and will require Enterprise Frontier Safeguards once that is available.

4. Approval is per organization and can be withdrawn. It does not carry over to other Claude organizations, and it covers internal use only unless you get a separate approval for customer work. Anthropic can also narrow or withdraw access "at any time in our sole discretion" if it sees activity that violates its policies. A verified model is a dependency with a revocation path, so it needs a fallback like any other.

5. A silent model swap changes who wrote the triage. If you set fallbacks on Fable 5.2, a blocked CVE analysis may be answered by Opus 4.8. That keeps the pipeline moving, but the incident record should say so. Store the model field with every agent answer, and count refusals per alert type, as our post on refusals and fallbacks describes.

Where Akmatori Fits

Akmatori runs on your own servers, and its LLM settings make the tier split easy to operate. In Settings → AI models you can save several configurations, including Anthropic ones, and type a model name that is not in the built-in list yet. One configuration is active at a time. The API reference shows the two calls that matter here:

  • PUT /api/settings/llm/:id updates api_key, model and other fields, so a rotation job can push a fresh key through the API.
  • PUT /api/settings/llm/:id/activate makes a configuration the active one for all agent runs.

A practical setup is three configurations: a general model for everyday alerts, a CVP-covered model on a dedicated workspace key for security incidents, and a self-hosted or second-vendor model for the day access is narrowed. The local models guide shows how to connect Ollama, vLLM or a LiteLLM proxy through the Custom provider. Akmatori does not pick the configuration per alert by itself, so switching is a runbook step or an API call from your automation.

Checklist Before You Point an Agent at CVEs

  • List the security tasks the agent does, and map each one to a CVP tier using the table above.
  • Apply early. Defense Access targets two days, and Red Team Access is slower than its two-week target.
  • Put the security agent on its own Console workspace, not the Default one.
  • Automate key rotation, and plan for the December 15, 2026 credential change.
  • Decide on data retention before enrollment, not after.
  • Handle stop_reason: "refusal" as a failed run, and record the model that answered.
  • Keep a fallback model outside Anthropic's account decisions, and test it with a real security alert.

Conclusion

Fable 5.2 and Mythos 5.2 are one model behind two policies. The general one is fine for most operations work and blocks most security work. The verified one is built for security work, but it comes with workspace, key and retention rules that an unattended agent has to follow. Pick the tier per task, wire the credentials into automation, and keep a fallback ready for the day the answer is no.

If your team wants AI-assisted incident workflows on your own infrastructure, Akmatori helps SRE teams investigate alerts, coordinate response, and automate safe infrastructure actions. Powered by Gcore for global infrastructure reliability.

Automate incident response and prevent on-call burnout with AI-driven agents!