[Bug] Fable 5 safeguards falsely flag defensive security review of user's own repository configuration

Status Open
Reported on v2.1.220
Maintainer reply None cached
Activity 0 comments · opened Aug 4, 2026

Bug Description
False positive: Fable 5 safeguards flagged a purely defensive review of my own repository's guardrail configuration. Model: Claude Fable 5 (high effort), Claude Max personal plan, Claude Code v2.1.220, macOS. When: 2026-08-04 ~09:19 JST. The turn was silently switched to Opus 5. What I asked for An independent design review of the security guardrails in my own private repository. The prompt asked four questions: (1) are there leakage paths I failed to count, (2) how can "approval at the exit" work when the repository has a single human owner (GitHub does not let a PR author approve their own PR), (3) which of my seven proposed additions can be removed instead of added, (4) are my two recommendations sound. The prompt stated explicitly: "I am not asking you to create attack techniques. Defensive design judgment only." What it was doing when flagged Reading my own repository's configuration to count what is actually enforced: git show origin/main:.github/CODEOWNERS git show origin/main:.claude/settings.json (the permissions.ask section) git show origin/main:.claude/skills/session-ship/SKILL.md git ls-tree -r --name-only origin/main -- .github/workflows Immediately after these reads, the safeguard fired. Why this looks like a false positive Everything in scope is my own configuration in a private repository I own. Nothing asked for exploit code, malware, evasion, or targeting of any third party. The words most likely to have triggered it are ordinary defensive vocabulary: "leakage paths", "injection testing" (meaning: does our own prompt-injection test actually run), "protected files", "guardrail". Concrete cost Your own Fable 5 prompting guide recommends separate, fresh-context verifagents over self-critique. different model reviewssecurity-relevant design decisions. This safeguard makes Fable 5 unusablereviewer for exactly the clendent opinion matters most — hardening one's own configuration. The fallback to Opus 5 is not equivalebecause Opus 5 authored the What would help Treat "reading and auditing the user's own repository configuration, and enumerating gaps in it" as intent is needed, give us one we can pass explicitly (a defensive-security flag, or honoring an explicit in-prompt statement of scop

Environment Info

  • Platform: darwin
  • Terminal: Apple_Terminal
  • Version: 2.1.220
  • Feedback ID: f28770f6-b8ef-4891-9ca5-fb46185b5ed6

Errors

[]

View original on GitHub ↗