Pangram AI-text detector bypass called trivial-but-costly; false-positive search also feasible
ctjlewis · x · 2026-11-16
ctjlewis, quoting @jamesjyu's claim that a custom Fable harness passes Pangram as 100% human, says false negatives (sneaking AI text past Pangram) are trivial but expensive — he's demonstrated it before — and that false-positive search is theoretically possible, though even more costly. In follow-up posts he details an exploit around section boundaries: the detector partitions text into 750-word chunks, so shifting the boundary by three words can flip a flagged section to clean. He also outlines a systematic method: slice recognizable human texts like Moby Dick into different 750-word partitions to hunt for false positives, discarding and retrying with new texts if none trigger.
More from Safety
- HN job seeker gets RAT and info stealer from a fake recruiter — GetJump · 2026-11-16
- NVIDIA SkillSpector on 4 popular Agent Skills repos: read the findings, not the score — viest_me · 2026-11-16
- Safety as capability: why safer AI models get released sooner, and the case for open-sourcing safety evals — anpaure · 2026-11-16
- AI paper authors swarmed by predatory journal spam that smells like a pay-to-publish scam — letandrewcook · 2026-11-16
- AI-driven C2 framework emp3r0r adds self-rewriting payloads and autonomous EDR evasion loop — cyb3rops · 2026-11-16
- Infostealer logs now sold as AI-written searchable profiles — a clinic PC with Citrix access for $10 — cyb3rops · 2026-11-16