anthropics / anthropics/claude-code

[Bug] Cybersecurity classifier blocking defensive malware analysis tooling mid-conversation

Ouverte
#88,014 0 commentaires 0 réactions 0 personnes assignées Voir sur GitHub
area:model area:security 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**
A conversation building defensive Android malware triage tooling (static APK analysis, permission auditing, LLM-assisted classification via Ollama Cloud, passive OSINT for IOC reporting to law enforcement) was blocked by the cybersecurity safety classifier partway through. The conversation had been proceeding normally for ~15 turns with no prior flags, including file creation of manifest audit code, router logic, and LLM client wrapper — all deterministic/defensive in nature.

Context:

Use case: triaging a real fake-KYC banking-dropper APK received via a compromised family member's WhatsApp, for submission to Maharashtra Cyber / CERT-In.
Explicit scope boundary was documented in the project README (passive analysis and OSINT only; active exploitation of attacker infrastructure explicitly excluded).
No malware was requested to be written; only detection/analysis tooling.
The block occurred on a message continuing prior technical work, not on a new/escalated request.

What I'd like investigated:

Whether this pattern (multi-turn defensive tooling requests, code generation for static analysis) is a known trigger for over-blocking.
Whether Cyber Verification Program enrollment status is correctly checked/applied mid-conversation, or only at session start.

Attachments available on request: conversation transcript, repo scaffold created prior to the block.

**Environment Info**
- Platform: darwin
- Terminal: iTerm.app
- Version: 2.1.212
- Feedback ID: 2fb7f73d-2023-4568-bb31-44c16e6f03ec

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

Guide de contribution

Aucun guide de contribution indexé pour ce dépôt

Piste de recherche

Start with the supplied conversation transcript and repo scaffold, especially the README scope boundary and the manifest audit, router, and LLM client wrapper mentioned in the report. Reproduce or review the block if the materials are provided, then determine whether defensive multi-turn requests trigger over-blocking and whether Cyber Verification Program enrollment is applied mid-conversation; done means both questions have a documented finding.

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

Évaluation

Stack technique
android, ollama, python
Domaine
ai, devtools, security
Type d'issue
Bug
Difficulté
4/5
Temps estimé
3-5 jours
Activité
Active
Clarté
À clarifier
Accessibilité débutants
35/100

Recevez les nouvelles issues par e-mail

Un résumé court des issues GitHub adaptées aux débutants.