OpenBles Firewall 1 拦截提示注入，对比 Jev 误报

Stanislav Sorokin @stas_sorokin_原推发布于 2026.10.10

在 X 上打开原推

Today we're releasing OpenBles Firewall 1: an open model that blocks prompt injection before your AI agent reads it. 31.8 MB. 33 ms on a laptop. The weights are yours, MIT. It is a decision model: it reads what your agent is about to read (a web page, an email, an MCP tool result) and blocks the text that tries to give it orders. → 119 of 120 attacks on agents caught, 0 clean texts flagged (public held-out set, 12 attack families) → 33 ms per text on a laptop. No API, no key, no bill per call → The weights, the training script and the eval are all in the repo, MIT → Trained in 8 minutes on a MacBook, on public human-labelled data The twist: on the same texts Jev caught 120 and flagged 11 clean ones. Ours flagged 0. Every agent that reads the open web needs a firewall in front of it. This one is free and runs on the same machine as the agent. What stands between your agents and the pages they read? Repost this now, because every team connecting MCP servers this week is feeding its agent text it did not write. Follow @stas_sorokin_ for open models and numbers from AI systems in real production. Weights in the reply.