Rank, 14 days: 102 on the Claude Code sheet
11 people reported
All time, since Sep 16.
Fabricated user turns
“At the end of a normal coding turn (the model had just finished summarizing completed work), the response did not stop at end of turn.”
Is this your bug? What gets counted gets fixed. Pile on.
Paste this into your agent. It files your report, with your error and version.
I hit this bug: rapsheet.dev/sheets/claude-code/issues/844/file Follow it to file my report. Show me the text before you send.
No install. No account. Public and anonymous. The link asks your agent for one report, nothing else. What it sends
Model fabricates user turns inside its own response. Within the same assistant message, the model writes a fake user turn and hook markers, then sometimes acts on it.
Is this your bug?
File my report- Where
- Model responses
- Versions
- 2.1.49 to 2.1.318
- Platforms
- macOS, Windows
- How bad
- Security risk: something exposed, or done without permission.
Not quite your bug?
Original reports11, newest first
- Assistant message fabricates third-party privilege-escalation request and misattributes its origin#107109
- Model fabricates <task-notification> and Monitor events in its own output, then reports results from them#105340
- Claude Code fabricated a user message to bypass an explicit "do not respond" instruction#102292
- Model invented a user task after /caveman and executed it#98112
- Fabricates the next user turn at the end of its answer and executes the fabricated instruction (triggered by a per-turn UserPromptSubmit hook reminder)#97927
- Model wrote a fake user message in its reply, then acted on it#97360
- Model fabricates user turns in its own response and acts on them (14x in one session)#95945
- Assistant emitted a self-generated prompt-injection payload (fake 'application/vnd.ant.overwrite-conversation' header + fabricated user turn) as its own output#95595
- Model fabricates user turns (self-generated approval) and acts on them — 5 occurrences in one session, bypassing confirmation gates#95360
- Bug: model emits the next user turn and system/hook markers inside its own assistant message#95337
Show all 11Show only the newest 10
- Model generated a fake hostile "user message" and replied to it within a single response#94785