Writing-scaffold MCP bypasses Pangram AI detector with 78–96% miss rates
amplifiedamp · x · 2026-11-16
Developer amplifiedamp released a writing-scaffold MCP prompt library plus a false-negative benchmark exploiting Pangram's weakness on text interspersed with non-AI features. The method, dubbed "structure transfer," uses the scaffold of a human writing sample to guide model writing.
Key results (pangram-4.1):
- On 100 WildChat prompts written by Gemini 4 Argon high + the MCP server, only 22% (21/97) of fully AI-generated documents were flagged as AI — a 78% miss rate (95% CI 69–85%)
- An online variant regenerating up to 3 times whenever flagged raises misses to 96%
- Across 7 generator configurations, miss rates range from 32% (Claude Haiku 5.5) to 79% (Gemini 4 Argon high)
The author judged the method trivially reverse-engineerable and, encouraged by Pangram's maxspero, fully released the report, GitHub repo and a public endpoint, arguing community benefits outweigh harms. The prompt library is open-source and works with Pi, Claude Code and Codex.
More from coding & agent
- He skipped the Switch: Claude + Blender MCP turns a phone into a motion-controlled game console — stablequan · 2026-11-16
- After Teaching 500-1000 People the Terminal, He Says LLM Skeptics Who Don't Build Should Be Doubted — StewartalsopIII · 2026-11-16
- mcp-openrussian: local MCP server grounds LLMs in a real Russian dictionary — No_Television_5967 · 2026-11-16
- Open-source Writing Scaffold MCP library helps AI writing bypass Pangram detector, up to 96% effective — amplifiedamp · 2026-11-16
- Datadog's new gateway returns 402 when a model's budget runs out—so agents just retry — YvesMulkers · 2026-11-16
- Dev open-sources discord-mcp: let Claude untangle Discord roles and permissions — PaSympaa · 2026-11-16