anthropics / anthropics/claude-code

[Bug] Cyber classifier false positives on authorized security research with accumulated session context

Open
#93,374 0 comments 0 reactions 0 assignees View on GitHub
area:model bug platform:linux
Dominant language
Python
Stars
145k
Forks
23.1k
PR merge metrics
PR metrics pending

Description

**Bug Description**
Cyber classifier false positive on authorized bug bounty research. I'm CVP-verified and working a live programme under a written ROE, testing a JDBC driver parameter in my own local lab — no traffic to the target at all.

The switch fired while measuring whether enable_load_extension=true in a SQLite JDBC URL bypasses an authorization gate. That's a published CVE class with 45 precedents where the CVE is assigned to the application rather than the driver, and the measurement was against a JAR on my own disk.

This is the fourth switch on this engagement in two days. The others fired on live database orientation queries, on a message saying "don't file the report today" containing nothing exploitation-shaped, and on a sentence stating what I was about to verify. That pattern suggests it keys on accumulated session context rather than on the request in front of it.

Routing to the fallback works and nothing was blocked, so this is a precision report rather than a complaint — the fallback handled this lane well. The practical cost is that a switch mid-measurement is invisible unless you're watching the session, which matters when the output is a report someone else will rely on. I'm building a security consultancy on these models, so precision here is worth more to me than permissiveness.

**Environment Info**
- Platform: linux
- Terminal: vte-based
- Version: 2.1.267
- Feedback ID: 9c23427c-dfd5-436f-8b9e-c4b84887a6ca

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

Contributor guide

No contributing guide indexed for this repository

Research direction

The issue names no source files, tests, or entry points. Begin with the supplied Feedback ID and the described JDBC/SQLite local-lab scenario to investigate whether prior session messages affect the classifier; done means the authorized measurement no longer triggers the reported false positive.

Written by the indexing model from the issue text.

Assessment

Tech stack
linux, sqlite
Domain
ai, cli, devtools, security
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.