Incident sheet · /security

Filed
Read
3 min · 664 words
Severity
Routine

Claude Fable 5.2 cuts false blocks; Mythos 5.2 finds more bugs

Anthropic says Claude Fable 5.2 cut unnecessary cyber blocks by about 60%, while Mythos 5.2 found 14.4% more vulnerabilities than its predecessor in testing.

Anthropic’s Claude Fable 5.2 blocks fewer security tasks, while Mythos gets better at finding bugs
FeedAnthropic’s Claude Fable 5.2 blocks fewer security tasks, while Mythos gets better at finding bugs — Innovalabs / pixabay

Event log

  1. Mythos 5.2 detects 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 on CVE Rediscovery.
  2. One in six confirmed Mythos 5.2 bugs was missed by every other frontier model in one security company's tests.
  3. Fable 5.2 reduced unnecessary cybersecurity blocks by about 60% compared with Fable 5.1.
  4. Mythos 5.2 reached selected partners, while Fable 5.2 reached regular users.
  5. High-risk Fable activity can switch automatically to Opus 4.8, and Mythos 5.2 underwent sandbox-escape testing.

By Olivia Hart

Anthropic’s Claude Mythos 5.2 detects 10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Mythos 5.1, according to the company’s CVE Rediscovery evaluation.

Anthropic recently rolled out Mythos 5.2 to selected partners and Claude Fable 5.2 to regular users, with improvements in coding and reasoning. The release also divides security work by audience: one model targets advanced cyber engineering, while the other keeps tighter controls for regular access.

Anthropic calls Mythos 5.2 its "most capable model for cybersecurity engineering" and says it now leads Fable 5.2 in security research. That claim concerns practical security workflows rather than general-purpose coding, where both releases received improvements.

Where Mythos 5.2 leads

Anthropic says Mythos 5.2 outperforms Fable 5.2 in three linked tasks:

  • Finding vulnerabilities.
  • Reproducing vulnerabilities.
  • Developing fixes.

The company designed Mythos for organizations handling advanced cybersecurity work. On CVE Rediscovery, it detected roughly one additional vulnerability for every 10 found by Opus 5.5, and one for every seven found by Mythos 5.1.

A security company that tested the system told Anthropic that "one in six confirmed bugs" identified by Mythos 5.2 had been missed by every other frontier model it tested. Put another way, about 17 of every 100 confirmed findings represented bugs the comparison models had overlooked.

Same model, different guardrails

Fable 5.2 and Mythos 5.2 share the same underlying parent model, but Anthropic applied different safeguards for their intended uses. Fable received additional cybersecurity and biology controls so it could serve regular users.

Those controls block most potentially dangerous hacking activities, including penetration testing, exploit generation and certain vulnerability-research tasks involving binary analysis. The restrictions remain narrower than the capabilities Anthropic grants through Mythos.

When Anthropic detects high-risk activity in Fable, it automatically switches cybersecurity-related responses to Opus 4.8. The company carried these restrictions into Fable 5.1, where even normal requests sometimes triggered blocks.

One example makes the problem concrete. A request to have Fable 5.1 log into an AWS droplet switched to Opus 4.8, which classified the activity as high risk. The block was described as unnecessary because Fable had originally configured the AWS account through an API.

Fewer false blocks

Anthropic says Fable 5.2 reduced unnecessary blocks by approximately 60% for cybersecurity requests compared with Fable 5.1. In practical terms, roughly three in five such blocks no longer occurred during early testing.

Anthropic made that observation across Claude.ai, Claude Code and Cowork. The improvement does not remove restrictions: Fable 5.2 still blocks high-risk work and can route cybersecurity questions to Opus 4.8.

Anthropic also tested Mythos 5.2 against sandbox-escape attempts, where an AI agent tries to access files or credentials outside its permitted environment. The company did not publish a pass rate or escape result. It acknowledged that its evaluations cannot catch every potential failure before release, particularly as models become more capable of recognizing when they are being tested.

What does this change for security teams?

Security teams with partner access can put Mythos 5.2 into vulnerability discovery, reproduction and remediation. Organizations using Fable 5.2 should expect routine cybersecurity work to run with fewer false stops, but high-risk requests may still be blocked or rerouted to Opus 4.8. The key distinction is access and control, not a wholesale removal of limits.

Separate product promotion

An accompanying advertisement promotes Adaptive Security’s AI agents, which analyze email intent, context and behavior to identify phishing, business email compromise and invoice fraud. It offers a free 14-day, read-only scan of Microsoft 365 or Google Workspace, with no MX changes. The promotion frames the scan as a way to uncover email threats existing defenses may miss.

For enterprise buyers, Mythos 5.2 brings stronger cyber tooling behind partner access, while Fable 5.2 offers broader availability with tighter guardrails.

via bleepingcomputer.com (Source)

Filed under

  • anthropic
  • claude
  • ai-security
  • vulnerability-detection
  • ai-guardrails
Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

News editor covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.

207 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›