[Bug] Anthropic API Error: False positive from the Fable 5 biology safeguard
Status Open
Reported on v2.1.222
Maintainer reply None cached
Activity 1 comment · opened Aug 5, 2026
Bug Description
see chat: Fable 5 session got flagged by the biology safeguard on Delphi-2M health-data content and switched to Opus 5 for good
Environment Info
- Platform: darwin
- Terminal: vscode
- Version: 2.1.222
- Feedback ID: 224fc037-a610-474e-9621-8858f826f6c2
Errors
[]
False positive from the Fable 5 biology safeguard,
plus a related model-switch behaviour issue.
WHAT I WAS DOING
Preparing a Norwegian-language briefing document for
a working session on Danish and Norwegian health-
registry data infrastructure. The flagged turn was
source-attribution QC: verifying citations in that
document and retrieving a published Nature paper to
check one of them.
WHAT APPEARS TO HAVE TRIGGERED IT
Descriptive text about a published, peer-reviewed
disease-progression model (Delphi-2M, Shmatko et al.,
Nature, 17 September 2025, doi 10.1038/s41586-025-
09529-3): that it was externally validated on 1.9
million people's registry data and models rates for
over 1000 diseases up to 20 years ahead. Alongside
this, citations of Danish health law (sundhedsloven
paragraph 46) on researcher access to journal data.
WHY THIS IS A FALSE POSITIVE
This is public-health policy and citation-verification
work built on an already-published Nature paper. The
content is epidemiological and legal. There is no
dual-use element, no protocol, and no capability
uplift of any kind. The model is being asked to check
whether a document cites its sources correctly.
SECOND ISSUE, POSSIBLY MORE ACTIONABLE
The switch to Opus 5 did not revert. After the flag at
15:20:51 UTC on 2026-08-05, all 26 subsequent
assistant messages ran on Opus 5, including across a
new user turn, with no return to Fable 5. Both the
in-product notice and my reading of the docs suggest
this is meant to be a per-message fallback. If that is
the intent, this looks like a bug. If it is meant to
be sticky for the session, the notice should say so,
because the practical effect is losing Fable 5 for the
remainder of a long working session after one flagged
message.
IMPACT
This is the second recurrence of this pattern for me.
Health-economics and health-data research is my
primary use case for Fable 5, and it is the use case
most reliably interrupted by the broad bio classifier.
Session: e7868abf-3628-4669-b94d-14d86309f0efThis issue has 1 comment on GitHub. Read the full discussion on GitHub ↗