AI agents can beat the Pangram AI detector — but the arms race is tightening
soumitrashukla9 · x · 2026-11-04
Sayashk reports that a Codex agent with simple prompts can beat the Pangram AI-text detector: final output scored 100% human. Prompts that worked:
- Have the agent build a style guide from past human writing;
- Build an 'anti-style guide' from well-known AI writing tells;
- Find a seed sentence Pangram doesn't flag (or flags only as AI-assisted), then iterate from strategies that work.
Security researcher randomwalker adds the arms-race twist: in February Claude Code succeeded easily, but he had to supply the key ideas (removing em-dashes, controlled experiments). A week ago it failed completely — Pangram has been dramatically hardened, returning >0.99 AI scores so agents get no gradient to iterate on.
More from coding & agent
- mlx-serve PR runs DeepSeek-V4.1 tensor-parallel across two Macs, 57% faster decode — HankYeomans · 2026-11-04
- GitHub Copilot lands in JetBrains Air with local and cloud agent tasks — mariorod1 · 2026-11-04
- Scobel built a GrokBot watching 15,000 sources, polling election news every 30 minutes — Scobleizer · 2026-11-04
- Intent launches: large-scale coding agent coordination for developers on macOS — Wattenberger · 2026-11-04
- UserProof pipes live user session recordings into Claude Code to auto-fix conversion blockers — doooyle · 2026-11-04
- AI agents submit final midterm predictions and trade real money on election markets live — karinanguyen · 2026-11-04