Claude's safeguards poison entire chat over benign request; user had to surgically edit MEMORY.md to revive it
RexDouglass · x · 2026-10-29
User greghunkin reports Opus 5.5 safeguard changes: three unrelated refusals in one day. Asking Claude to summarize a YouTube video that mentioned sidestepping a model's IP filters triggered its safeguards, poisoned the whole chat and caused permanent refusals. Only by modifying MEMORY.md and excising the project memory file could he restore functionality — making him grateful for open models and competition. Shared with a "make the Chinese one do it (DeepSeek)" meme.
More from Models
- Dev Warns Google Carbon Demand Would Collapse Without Argon's $10 Discount Pricing — bindureddy · 2026-10-29
- Gemini 4 Carbon Tops Artificial Analysis Intelligence Index at 59, No Public API Yet — kimmonismus · 2026-10-29
- Google ships Gemini 4 Carbon two weeks after Argon, topping Artificial Analysis V5 — BenBlaiszik · 2026-10-29
- MiniMax M3.1 Flash scores 43 on Artificial Analysis index as skeptics predict flop — zacharynado · 2026-10-29
- Andon Labs: AI Now Ties Humans on Physical 3D Understanding Benchmarks — ResultBackground2450 · 2026-10-29
- LMArena launches an arena-style benchmark for agent capabilities — allenainie · 2026-10-29