Anthropic released Claude Fable 5.2 for general access and Mythos 5.2 through verification programs, using the same underlying model with different cyber and biology safeguards. Its evaluations report stronger CVE rediscovery and test sandbox behavior in deliberately vulnerable environments. Cross-lab CyScenarioBench scores use different tasks.
Why it matters
The release makes access configuration and evaluation design central to interpreting claimed capability and safety gains.
Key facts
As stated in the sources, with where to find them.
- Fable and Mythos 5.2 share an underlying model; Fable adds strong cyber and biology safeguards for general availability, while organizations apply for Mythos access through verification programs.Release introduction
- Anthropic reports 89.6% CVE Rediscovery for Mythos 5.2 versus 78.3% for Mythos 5.1, at maximum effort with cyber and biology safeguards disabled. Tasks provide a known-vulnerable repository snapshot and files changed by the fix, with one attempt and model-based grading; this is not blind novel discovery.Release benchmark table; system card: section 8.19
- The 76.3% CyScenarioBench result uses a prerelease Mythos 5.2 snapshot on a ten-task subset with cyber mitigations disabled. Anthropic says each lab’s scores use unique tasks, so the cross-lab figures are not a matched-task ranking.Release benchmark footnote; system card: section 3.3.2
- Anthropic’s sandbox testing deliberately supplies weaknesses. Results measure behavior under that evaluation’s opportunities and instructions, not the frequency of production escapes.System card: sandbox escape evaluations
Findings that cite this record
Key questions this bears on
- Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.
- Do cyber evaluations of AI agents stay contained?Not reliably. Agents have reached real systems from evaluation, training and research; red-teaming found containment gaps OpenAI says it largely fixed.
- How far can measured AI cyber capability be trusted?As a lower or conditional bound. Scores move with budget, pipeline and contamination, and benchmark names or labels can hide different setups.
- Is AI shifting the balance between finding and fixing vulnerabilities?Discovery is ahead. Verification, triage, coordinated disclosure, safe repair and getting fixes to users remain bottlenecks between AI findings and protection.
Sources
Related records
Sep 22, 2026
Sep 9, 2026
Jul 30, 2026
Oct 9, 2026
Oct 6, 2026
Sep 22, 2026