Anthropic's harness blocks harmless git commits but silently trims context and uploads drafts
maier_ak · x · 2026-09-15
A detailed writeup documents three concrete Claude harness problems: the git commit permission toggle flips from green to red mid-session with no warning and can't be re-enabled; context gets silently compressed well before the advertised limit, orphaning commands and losing execution links; and half-finished Artifacts drafts get uploaded to the cloud without user awareness. The author's verdict: the harness audits the fire extinguisher (harmless actions) while ignoring the gas leak (private data leaking).
More from Models
- Coverage tests across 3 runs: DeepSeek V4.1 Flash 21.7→33, GLM-5.3 Flash 17.7→24 — PawelHuryn · 2026-09-15
- ChatGPT Plus vs Business: heavy users compare real-world Sol High usage limits — geler1 · 2026-09-15
- Tinfoil: Open-source safety classifiers are scarce, frontier labs urged to release more — yuntiandeng · 2026-09-15
- Depending on a single AI vendor is riskier than open models, argues researcher — xeophon · 2026-09-15
- Grok 5 Roadmap, Gemini 4.0 Pro 'Argon' Leaks, and Claude Opus 5.2 Preview Roundup — WorldofAI · 2026-09-15
- User verdict after testing small models: DeepSeek v4.1 Flash is the minimum viable model for real work — solyarisoftware · 2026-09-15