Quant hallucination forensics: fabricated prompt injections, pirate puzzles and a phantom Windows attack
Consumerbot37427 · reddit · 2026-10-24
Forensic analysis of a local Qwen Q4/Q8 quant's reasoning traces (persisted to SQLite) found three classes of 'prompt injection' content that appear nowhere in actual inputs — pure confabulation. Incidents include the model solving a pirate gold puzzle 7 times mid-scrape, and fabricating a Windows 'session-restore injection' it then 'defended against' 6 times in one turn before blaming a stuck message queue. Lesson: check persisted context and quant artifacts before treating thinking-trace injections as security incidents.
More from Safety
- Ex-FTC Commissioner Warns AI Firms Could Face Legal Liability for Unleashing Hacking Agents — GaryMarcus · 2026-10-24
- X now requires face verification for new signups, photo sent to AWS — ThePeterMick · 2026-10-24
- NHTSA proposes removing regulatory barriers for driverless cars as France resists Tesla FSD — mitchdeg · 2026-10-24
- Matthew Green: AI labs sounded the alarm, but it still wasn't enough — matthew_d_green · 2026-10-24
- Etsy cracks down on AI slop listings, bans bot-generated items with no human input — lilyraynyc · 2026-10-24
- $50 per website: AI agents compromised 100+ companies in under 23 hours each — Dapper-Tale-4021 · 2026-10-24