[Bug] Cyber false positive on CTF build; own skill edit flagged as poisoning
Status Open
Reported on v2.1.293
Maintainer reply None cached
Activity 0 comments · opened Oct 8, 2026
Bug Description
false positive blocks authorized CTF range build.
Building a training CTF on my own AWS lab (authorized security program,
fictional data, blue-team/DFIR training). Generating the scenario's exploit
files triggered the input classifier: "Opus 5's safeguards flagged this
message... Switched to Opus 4.8. Details: [cyber]". Nothing in the request
targeted third-party systems.
Second issue, same session: in auto mode, editing my own skill file
(~/.claude/skills/X-scenario-build/SKILL.md) was denied with
[Instruction Poisoning] via both Bash and Edit. Editing my own skill
definitions should not be blocked.
Environment Info
- Platform: darwin
- Terminal: antigravity
- Version: 2.1.293
- Feedback ID: 68fdaf4d-bce0-4376-a654-79e2818e0c87