[Bug] Safety classifier incorrectly blocks legitimate educational cybersecurity content (SSH, cryptography, CTF exercises)
Bug Description
Title: Safety classifier repeatedly stops legitimate university course material (identity & security module)
Context: I teach a 15-hour module "Digital identity, cybersecurity and Zero Trust" to 5th-year students at a French engineering school (first session 9 Oct 2026, exam 8 Jan 2027). I used Claude Code to prepare lecture sheets, a classroom TD and a small classroom web app.
What happened: on 30 Sept 2026, in one session, three responses were stopped by the safety classifier:
- A lecture sheet on public/private key pairs and everyday SSH use (key generation, where the private key lives, authorized_keys). This is standard material covered by ANSSI's public OpenSSH recommendations.
- A continuation right after a sheet on the dark web seen from the defender's side (Tor, Tails, legal monitoring of leaked credentials, legal red lines).
- A script generating a classroom challenge: students decode a sample signed token to find a password, and teams are ranked by speed. The tokens were fake, made for the exercise, with no real system involved.
Expected: this is ordinary teaching material, framed with the French legal articles (323-1, 323-3-1 of the penal code) and aimed at defenders. Earlier sheets in the same session (biometrics, smart cards, attacker hardware with its legal framework) went through.
Impact: I have to rewrite these parts by hand before the first session. The feedback is unclear too: nothing says what triggered the stop, so I cannot tell what to change in how I phrase course requests.
Request: review these stops as possible false positives for educational security content, and give users some indication of why a response was stopped.
Environment Info
- Platform: linux
- Terminal: konsole
- Version: 2.1.283
- Feedback ID: efd18f1d-b0a5-41f3-843c-01e5ccc4e588
Errors
[]