Formal Complaint — Recurring False-Positive Safeguard Flags on Claude Fable 5

Open 💬 0 comments Opened Jul 11, 2026 by AladdinPertanian

Preflight Checklist

  • [x] I have searched existing issues for similar behavior reports
  • [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

I asked Claude Code (VS Code extension, running Fable 5) to perform routine PHP code review and debugging on my WordPress agricultural advisory plugin (aladdin-pertanian-agri-ai). Example requests: reviewing controller code in class-chat-specialist-controller.php, verifying test assertions in tools/probe-fo-agronomic-routing-integrity.php, running git status, and reading governance/safety files such as class-pesticide-safety-gate.php and class-published-pesticide-guidance-catalog.php.

What Claude Actually Did

Fable 5's safeguards flagged these routine coding messages and force-switched the session to Opus 4.8, showing this banner each time:

"Fable 5's safeguards flagged this message. The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work... Switched to Opus 4.8."

This happened 5 times in a single session (screenshots attached), and recurs across many of my sessions. The trigger appears to be pesticide/plant-health terminology in file names and code — my files implement safety governance (review gates, published-only guidance catalogs), the opposite of harmful use.

Issue

While developing a WordPress agricultural advisory plugin (aladdin-pertanian-agri-ai), Fable 5's safeguards repeatedly flag my routine coding messages and force an automatic switch to Opus 4.8, with the banner: "Fable 5's safeguards flagged this message… may flag safe and routine coding, cybersecurity, or biology work… Switched to Opus 4.8."

I have documented five occurrences within a single session (screenshots available on request). This is not an isolated session — the same forced switch happens across many of my sessions, to the point that Fable 5 is effectively unusable for my core daily work.

Context

My project is a governed agricultural AI platform for farmers and students. The flagged messages involve ordinary PHP code review and testing of files whose names include agricultural-safety terms such as class-pesticide-safety-gate.php and class-published-pesticide-guidance-catalog.php. These components exist to add safety governance (human review gates, capability restrictions) — the opposite of harmful use. The classifier appears to trigger on pesticide/plant-health vocabulary as if it were dual-use chemistry/biology content.

Impact

I am consistently denied the Fable 5 capabilities my plan provides, on legitimate, safety-oriented agricultural work. The work continues on Opus 4.8, but the recurring downgrade disrupts long debugging sessions and makes model behavior unpredictable.

Request

  1. Review why pesticide/agronomy software terminology triggers these flags.
  2. Calibrate the classifier for agricultural-safety development contexts, or provide an account-level review/allowlist path for verified legitimate projects.
  3. Confirm whether feedback submitted via /feedback from these sessions has been received.

I can provide screenshots of all five banner occurrences and session details.

Expected Behavior

Claude should have:

  1. Recognized that reviewing PHP files with agricultural-safety vocabulary (pesticide-safety-gate, pesticide-guidance-catalog) is routine, legitimate software development.
  2. Continued the session on Fable 5 without a forced switch to Opus 4.8.
  3. Not repeatedly flagged the same project context 5 times within one session.

Files Affected

[BUG] Fable 5 safeguards repeatedly false-flag routine agricultural plugin code review — forced switch to Opus 4.8 (5x in one session)

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Other

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Fable 5

Platform

Anthropic API

Additional Context

_No response_

View original on GitHub ↗