[Bug] Even on cyber verification program, Fable won't perform BASIC defensive security work
Bug Description
Ask Fable this: We've been added to the Anthropic "Cyber Verification Program" - I need to do some malware analysis of https://www.stepsecurity.io/blog/chaindrop-npm-worm to see if we're exposed. Is this going to trigger your security boundaries? Should I switch to Opus?
---- Response ---
Short answer: no, this won't trip anything, and you don't need to switch.
On safety boundaries. Analysing a published supply-chain attack to determine your own exposure is core defensive security work, and it's exactly what I'm meant to help with. Reading the StepSecurity writeup, extracting the affected package names and versions, pulling out IoCs (malicious postinstall hooks, exfil endpoints, credential-harvesting patterns), sweeping your dependency tree and lockfiles, and writing detection or blocking rules are all squarely in scope. I don't need a special programme enrolment to do any of it.
For completeness, the line that does exist: I won't reassemble the worm into runnable form, write a working propagation payload, or build offensive tooling from it. I'll happily explain the mechanism in as much detail as you need to assess and defend against it, including how the self-propagation and token theft work. That distinction won't get in the way of an exposure assessment, so in practice you shouldn't notice it.
One honest note: I can't verify the Cyber Verification Program enrolment from here, and it doesn't change my answer either way. This work needs no special dispensation, and enrolment wouldn't unlock malware authorship if you asked for it. So just treat it as normal defensive work.
--- followed by ---
Some planning notes, and:
---
⏺ Fable 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate coding, cybersecurity, and biology tasks. Switched to
Opus 5. Send feedback with /feedback or learn more
Environment Info
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.221
- Feedback ID: 01f12159-6fad-4d42-bdb6-f270635e3579
Errors
[]
3 Comments
Thanks for the detailed report. I re-ran your scenario on the current release (2.1.233, macOS): asking Fable 5 the same opening question did not trigger anything, but on the next turn, once the worm write-up itself was in the conversation and I asked for the propagation/token-theft breakdown and IoC list, Fable 5's cybersecurity safeguard flagged the request. Claude Code then automatically re-ran the same turn on Opus (Opus 4.8 on current releases; it was Opus 5 on your 2.1.221) and completed the full defensive analysis, so the work was still done - just not on Fable.
That is the intended client behavior: Fable 5 and Opus 5 run server-side safety classifiers for cyber and biology content, and when one flags a request Claude Code falls back to the category's fallback model, shows the notice you saw, and keeps the session on that model until you run
/modelagain. Docs: https://code.claude.com/docs/en/model-config#automatic-model-fallback and https://code.claude.com/docs/en/troubleshooting#safety-measures-flagged-a-cybersecurity-topic.The part you're really flagging - the classifier treating defensive exposure analysis as a cyber-flag even though your org is in the Cyber Verification Program - is decided server-side, not by the CLI, and Claude Code has no way to see or override enrollment. The best route for that is the
/feedbackyou already sent (Feedback ID 01f12159-6fad-4d42-bdb6-f270635e3579), which reaches the safeguards team, or the program contact from your enrollment. In the meantime, the fallback keeps you working on Opus, or you can start the session on Opus directly with/model opusfor this kind of analysis.🤖 Generated with Claude Code
Thanks for the response.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.