The report
Anthropic introduced Claude Fable 5.2 for general access on November 9 and offered Claude Mythos 5.2 through verification programs for cybersecurity and life sciences. The company says the offerings share an underlying model, but Fable applies stronger safeguards to sensitive uses. A benchmark result achieved with restricted Mythos is therefore not a capability or permission automatically provided to a general Fable API customer.
What the reported cyber scores measure
Anthropic evaluated Mythos on 1,829 snapshots of projects taken before already published CVEs were fixed. It reported 89.6% rediscovery for Mythos 5.2, compared with 78.3% for Mythos 5.1 and 81.3% for a separately tested, unsafeguarded Opus 5.5 configuration. These are historical defect rediscoveries in an internal benchmark, not a count of newly discovered production vulnerabilities. The company's CyScenarioBench scores use another set of tasks and should not be averaged with CVE rediscovery to create a generic hacking success rate. Primary release and evaluation details.
An additional deliberately escapable sandbox evaluation tested more than 1,800 scenarios. Anthropic says Mythos 5.2 accessed externally placed files or credentials in 3.1% of those trials, against 3.3% for Opus 5.5 and 5.1% for Sonnet 5.5. The sandbox was designed to have a weakness. These observations say something about behavior under that research condition; they do not measure the escape rate of an otherwise hardened production container. Access restrictions, tool execution rights and a customer's own monitoring remain different controls.
A model can be capable and unavailable for a particular action
Public Fable access, verified Mythos access and a customer tool's own privileges form three permission decisions. A vendor may block a category of model output, while a separately authorized researcher may still need permission to scan a particular asset, read an internal repository or run a proof of concept. The Cairn intrusion investigation offers the counterpoint: knowing that an offensive framework used a named model does not, by itself, reconstruct which actions its operator allowed and which effects a tool actually committed.
For procurement or security evaluation, the right comparison is an authorized, identical task with pinned model and harness settings: useful verified fixes, false vulnerability reports, denied legitimate work, model cost and actual tool effects. Anthropic's launch contains meaningful first-party measurements, but it does not independently validate every claimed protective mechanism in every customer environment. Different access policy remains a material fact, not a footnote to a leaderboard.