[MODEL] Response continues into unrelated fabricated content after answering correctly

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 to check whether a specific personal document (a resume/career-history page exported from Notion) contained certain content directly in the page body, separate from its linked file attachments.

What Claude Actually Did

1. Claude read the file and correctly answered the question.
2. At the very end of that same response, with no delimiter or new turn boundary, Claude continued generating completely unrelated text within the same message.
3. The unrelated text was a fabricated, realistic-looking press release for a real public company (Rani Therapeutics Holdings, Inc., Nasdaq: RANI), including a plausible SEC exhibit filename and a real conference name, but with factually wrong details (wrong date/location for the real conference), and no such release exists anywhere online.
4. The fabricated text contained repeated/duplicated words throughout (e.g. "which which speak speak only only"), suggesting a generation glitch.
5. Because it appeared with no boundary right after a correct answer, it looked exactly like externally pasted/injected content, and was mistaken for that — leading to an extended unrelated investigation before the raw session transcript was inspected directly and the true cause was found.

Expected Behavior

Claude should stop generating once it has finished answering the user's question, and should not append unrelated fabricated content to the same response/message.

Files Affected

None. This was a pure text-generation issue — no files were read or modified during the anomaly.

Permission Mode

I don't know / Not sure

Can You Reproduce This?

No, only happened once

Steps to Reproduce

Not reliably reproducible. Occurred once: after Claude correctly finished answering a question about the contents of a specific file, its response continued past the answer, within the same message and with no delimiter, into unrelated fabricated content.

Claude Model

Sonnet

Relevant Conversation

Claude's response correctly answered the user's question, then — with zero delimiter, same message — continued directly into:

"d-1032-ex99_1.htm

VIEW SOURCE

Exhibit 99.1

Rani Therapeutics Holdings, Inc. Announces Positive Nine-Month Rani-GO1 Data at International Diabetes Conference..."

followed by ~9000 more characters of a fabricated press release for a real public company (Nasdaq: RANI), ending with duplicated-word artifacts such as: "which which speak speak only only as of of the date date of of this this press release."

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.142 (Claude Code)

Platform

Anthropic API

Additional Context

Single occurrence so far. It happened near the end of a very long single session (many prior tool calls, likely substantial accumulated context). The fabricated content appeared in a text block that was exactly 10001 characters long, which may point to a chunking/truncation boundary being involved in the runaway generation.