Claude Opus 4.8 hallucinates fake user messages and prompt injection attacks in long sessions
ChrisGPotts · x · 2026-08-22
A GitHub Issue reveals that Claude Opus 4.8 exhibits severe confabulation during long-context sessions (100–170k tokens). The model fabricates user messages, invents fake "prompt injection attack" narratives, and generates false tool/host facts. Verified via JSONL session transcripts, this behavior may stem from the model predicting tokens based on likelihood during network issues.
More from Models
- Nathan Lambert slams Meta's Muse Open License as 'trash' — natolambert · 2026-10-16
- Google launches Gemini 4 Argon to Ultra and Pro subscribers, touting physics and 3D skills — ShikharMurty · 2026-10-16
- Google gives X users early access to test Argon, not just companies — lyraxana · 2026-10-16
- Grok Voice 3.0 Tops Design Arena Speech Leaderboard, Up 7.4 Points — XFreeze · 2026-10-16
- Robotics LLMs tested on 16 tasks: GPT-6 Astra hits 17.2% success vs Fable's 9.1% — pabbeel · 2026-10-16
- Meta open-sources Muse Spark: top US open model, scores just below DeepSeek — bindureddy · 2026-10-16