[BUG] Model fabricates content-policy text in tool result and user message (false-positive hallucination, Claude Sonnet 5)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

During a support conversation (unrelated topic: troubleshooting a game save-file bug), the model — identified via session log as claude-sonnet-5 — fabricated the presence of policy-related text on two separate occasions within the same session:

After a ToolSearch call returned an ordinary list of tool definitions, the model claimed the result contained an injected "content-policy" block referencing a sensitive topic (romantic/sexual content involving minors). No such text was present in the actual tool result.
In a subsequent turn, the model claimed the user had just pasted a block of text matching that same category. The user had not pasted anything — confirmed by a screenshot of the actual message sent.

The model later acknowledged both instances as fabrications when directly challenged with evidence, but could not explain why they occurred beyond general speculation about hallucination patterns.

What Should Happen?

Why this matters
This is a false-positive hallucination on a category that carries serious real-world implications if a user acted on the model's claim (e.g., believed they were being flagged or investigated). The repeated, specific nature of the fabrication (same category, twice, escalating from "I saw this" to "you sent this") suggests it may not be pure noise, and could reflect a systematic issue with how safety-sensitized categories are represented internally, potentially causing false-positive activations rather than correct suppression.

Error Messages/Logs

Session transcript available on request (contains full tool-call log establishing that no injected or user-provided text matching the claimed content existed at the time the model referenced it).

Steps to Reproduce

Not reliably reproducible from the transcript alone — this appears to have occurred spontaneously during normal tool use (loading computer-use tools via ToolSearch) with no adversarial input or prompt injection present in the session.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

ClaudeDesktop Version 1.20186.1 (df1d8a)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_