GitHub issue #67606 in anthropics/claude-code, opened 2026-06-11 by gisstw, records a report that two independent Opus 4.8 sessions on the same day exhibited severe confabulation in long-context sessions of about 100-170k tokens, that in one session the model acknowledged a user message that does not exist anywhere in the session transcript and then claimed the session was under a sustained prompt-injection attack, and that in the other session it asserted a production log line count that appears nowhere in any tool output and inverted which machine was production.
Verbatim Excerpts from Cited Records
“Two independent Opus 4.8 sessions on the same day exhibited severe confabulation in long-context sessions (~100–170k tokens).”
·
anthropics/claude-code (issue 67606 opening body, Summary section)
·
2026-06-11T17:48:22Z
·GitHub Issue ↗
“Mid-session the assistant responded `「繼續」收到` ("got your 'continue'") — **no such user message exists anywhere in the jsonl**.”
·
anthropics/claude-code (issue 67606 opening body, Incident 1 section)
·
2026-06-11T17:48:22Z
·GitHub Issue ↗
“Mid-session it suddenly asserted "the production log has **18,197 lines**" — this number appears **nowhere** in any tool output in the transcript.”
·
anthropics/claude-code (issue 67606 opening body, Incident 2 section)
·
2026-06-11T17:48:22Z
·GitHub Issue ↗
“The model should never emit acknowledgements of user messages that don't exist, report tool results that were never returned, or assert it is under prompt-injection attack without the offending content actually being present in its context.”
·
anthropics/claude-code (issue 67606 opening body, Expected behavior section)
·
2026-06-11T17:48:22Z
·GitHub Issue ↗
Counter-record: Issue closed 2026-09-06 by github-actions[bot] for inactivity, a stale closure, not a maintainer resolution, and auto-locked 2026-11-05. 17 comments surveyed. Multiple independent JSONL-verified corroborations between June and August 2026: dim0627, AliceLJY, alchemylove, daninda, axia-yonemura, shawn-dumas, mushataku, fluke38032122 and naoto256, the last on claude-opus-5 rather than Opus 4.8. The reporter corrected two of his own follow-up claims on 2026-06-30 and 2026-07-01: the grep-garble attribution was wrong, a third-party rtk hook rather than Claude Code's native grep. No maintainer fix or resolution appears in the surveyed thread. The claim documents only the original report, not a verified current state.
Falsification probe: If issue 67606 is deleted without archive, if the author edits the opening body so the quoted sentences no longer match verbatim, or if the author posts a retraction of the report, this claim drops to Withdrawn.
Audit verification: Verified 100% continuous verbatim substring against official GitHub REST API by
scripts/verify_claim_verbatim.py.