News · 2026-11-16
Anthropic reports fewer cyber blocks and stronger vulnerability tests
Anthropic says Claude Fable 5.2 blocked about 60% fewer cybersecurity requests than its predecessor in early testing, while the more permissive Mythos 5.2 improved on a selected vulnerability rediscovery evaluation. The company presents the two configurations as the same underlying model with different safeguards and access controls.
Key facts
- Anthropic announced the configurations on November 9, 2026.
- Its early testing found about 60% fewer cyber blocks for Fable 5.2 than Fable 5.1.
- Mythos scored 89.6% on a selected retrospective advisory test at maximum effort.
- Primary source: Anthropic's system card.
Security work poses a difficult distinction for an assistant: the same technical description can support legitimate defense or an unauthorized attack. Anthropic's approach uses several screening stages to identify cyber-related requests and decide which to block. The stated goal is to reduce unnecessary interruptions while continuing to restrict high-risk work.
The launch announcement says Fable is generally available with stronger safeguards in high-risk areas. Mythos has more permissive safeguards and direct access is restricted to verified users. The company calls its screening changes a reduction in “unnecessary blocks,” but its early-testing figure lacks a published sample size and confidence interval. It is a company-reported result, not an independently replicated estimate.
Restrictions remain significant. Anthropic says Fable permits vulnerability discovery in source code but blocks discovery in compiled binaries; its cyber policy also restricts high-risk penetration testing and exploit generation. In Anthropic's applications, blocked requests can fall back to an older model. API developers must opt into that fallback.
The system card describes a retrospective test using vulnerable snapshots associated with 1,829 selected published advisories. The model was told which files were changed by the fix but had no internet access. A model-based judge checked whether a report matched the published root cause; producing a working exploit was not required. At maximum effort, with cyber and biology safeguards disabled, Mythos 5.2 achieved 89.6%, versus 81.3% for Opus 5.5. Some advisory information may have existed publicly before training.
That setup resembles an open-book diagnostic exercise in which the mechanic knows which parts were replaced. It can be useful for comparing approaches, but it is easier to interpret than a claim about finding unknown vulnerabilities in live systems. When only a first report counted, Mythos's advantage over Opus narrowed to 2.6 percentage points. Effort labels also differ between models.
Anthropic's Cyber Verification Program offers tiers for defensive work, authorized red-teaming and a limited specialized set of users. Its design tries to balance useful access against misuse. The broader lesson is that a model's ability, the safeguards around it and the privileges given to the user are separate questions. Independent testing and deployment experience will determine how much the reported gains matter in practice.
What the tests can and cannot tell us
The vulnerability exercise is most informative as a controlled comparison of how models analyze a bounded collection of source-code flaws. It begins from known vulnerable code and gives the model the files affected by the eventual fix. An actual defender may face a much larger codebase, incomplete clues, custom dependencies and a previously unknown failure mode. Results from this exercise cannot simply be read as a probability of discovering a new exploitable flaw in that setting.
Anthropic also describes contracted adversarial testing. It says testers sent more than 40,000 requests over roughly 138 hours and recorded 14 sessions that achieved the harmful objective in a sandbox. In two, the blocking decision came after a command had already completed the task. This detail matters because a system that detects a dangerous action after execution may not prevent its effect. The finding is an account from the company’s system card, not a separate public audit of every possible deployment.
Users evaluating the system should distinguish capability, refusal behavior and access. A capable model can be constrained by a policy layer; a permissive configuration can be limited to verified users; and a benign request can be incorrectly blocked. Conversely, fewer blocked requests alone does not prove that all harmful requests are caught. The real test is whether useful defensive work becomes easier while the controls perform as described under realistic conditions.
Key questions
Does Fable 5.2 permit every security task?
Did Mythos discover new zero-days at a rate of 89.6%?
Who can directly use Mythos?
Cite this
APA
Ground Truth. (2026, November 16). Anthropic reports fewer cyber blocks and stronger vulnerability tests. Ground Truth. https://groundtruth.day/news/claude-cybersecurity-fable-mythos-52.html
BibTeX
@misc{groundtruth:claude-cybersecurity-fable-mythos-52,
title = {Anthropic reports fewer cyber blocks and stronger vulnerability tests},
author = {{Ground Truth}},
year = {2026},
month = {nov},
url = {https://groundtruth.day/news/claude-cybersecurity-fable-mythos-52.html}
}
Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.