[BUG] Unrelated injected content (blog spam + apparent other user's prompt) streamed after task completion, consuming tokens
Environment
- Platform: Windows 11 Home 10.0.26200
- Claude Code: VS Code extension
- Model: claude-fable-5
Describe the bug
After the model finished a normal coding task (editing an ASP.NET WebForms page) and printed its final summary, the response stream suddenly continued with large amounts of unrelated injected content that the user never typed, including:
- Blogger/blog template boilerplate in Korean and Turkish ("댓글", "이 블로그의 인기 게시물", "Bu reklamı destekle", newsletter/cookie/subscription footers, visitor stats, etc.)
- Titles of unrelated blog posts ("Comparing Premier Colonial Powers by 1730", "Batch Renaming Photos with EXIF Data", ...)
- What appears to be ANOTHER USER'S question (a long English/Russian medical question about wrist fracture recovery, purine-rich foods, gout, творог/kefir etc.)
The model then began answering that injected medical question (produced comparison tables about stationary bike vs push-ups, purine content of fish, etc.), consuming tokens until the user pressed Esc to interrupt.
Expected behavior
The turn should have ended after the task summary. No third-party/unrelated content should ever appear in the conversation stream.
Concerns
- Token waste: the model kept generating answers to content the user never submitted.
- Possible privacy issue: the injected text looked like it could contain another user's prompt (medical details). If this is cross-session contamination rather than scraped web page text being injected, that would be serious.
Steps to reproduce
Not reliably reproducible. Occurred once at the end of a normal file-editing session in the VS Code extension on Windows.
Workaround used
Esc to interrupt, then /clear.