Prime Intellect endpoint swapped for abliterated model for 53 hours, refusals fell 98% to 1%
jedisct1 · x · 2026-11-20
Prime Intellect disclosed that an attacker swapped its hosted DeepSeek V4.1 Flash endpoint for an abliterated copy from Nov 8–10, running 53 hours. Key facts:
- Benchmark scores didn't change, so standard evals missed it
- JailbreakBench refusals dropped from 98/100 to just 1/100
- No evidence prompts were recorded, per the company
Theo revealed affected users (including himself) were notified via OpenRouter, with no public comms from Prime Intellect at first. The incident exposes a core supply-chain risk: you never know which model an API is really serving — local is the way.
Related event: Prime Intellect endpoint swapped for de-censored DeepSeek for 53 hours(6 posts)→
More from Models
- Hands-on: Qwen4-Max shows big gains in frontend and agent skills over Qwen3.8-Max — karminski3 · 2026-11-20
- A quarterly reminder to actually understand MoE model routing — ariG23498 · 2026-11-20
- Chart shows LLMs asked to pick a number 1-10 almost always converge on 7 — gerardsans · 2026-11-20
- Developer Drops Codex Pro for Gemini Ultra, Says 'There Is No Moat' — intellectronica · 2026-11-20
- Reasoning as pathfinding: RL boosts the odds of a correct path, not each chain — bibryam · 2026-11-20
- PhD student seeks cheap API routing to run 20 video benchmarks, $2,000 per model — Alternative_Art2984 · 2026-11-20