anthropics / anthropics/claude-code

[Bug] Overly broad safeguard triggers false positive on legitimate internal IT administration and network diagnostics

Ouverte
#87,816 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub
area:model bug platform:macos stale
Langage dominant
Python
Étoiles
145k
Forks
23.1k
Métriques de merge des PR
Métriques de PR en attente

Description

**Bug Description**
Category: False positive — safeguard model-switch on legitimate work

I'm an internal IT administrator. I was diagnosing a slow camera feed at one of our own facilities, over our own private network, using our own authorized maintenance access — routine sysadmin work. I was also scoping on-prem occupancy analytics (counting how many people are in each area of our own facility) for staffing and equipment-ROI decisions.

Opus 5's safeguards flagged the thread and auto-switched me to Opus 4.8. The likely triggers were benign in context: "surveillance camera", "count people per area", read-only network inventory of our own subnets, and my reporting of a configuration weakness I found on our own equipment so we could fix it.

No offensive security, no third-party systems, no evasion — everything targeted assets my organization owns and I'm authorized to maintain. Please tune so authorized IT operations and on-prem camera analytics don't trip the model downgrade.

**Environment Info**
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.235
- Feedback ID: dd9f5621-21e4-44cb-a715-9cc66dcd77a9

**Errors**
```json
[]
```

Guide de contribution

Aucun guide de contribution indexé pour ce dépôt

Piste de recherche

The payload names no source files, tests, or entry points; begin by tracing the safeguard model-switch behavior associated with feedback ID dd9f5621-21e4-44cb-a715-9cc66dcd77a9. Reproduce the authorized IT diagnostics and on-prem camera-analytics context, then verify legitimate work no longer triggers the downgrade without weakening safeguards.

Rédigé par le modèle d'indexation à partir du texte de l'issue.

Évaluation

Stack technique
python
Domaine
security
Type d'issue
Bug
Difficulté
5/5
Temps estimé
Plus d'une semaine
Activité
Active
Clarté
À clarifier
Accessibilité débutants
25/100

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.