aiminute. ← All AI news
New Models AI Minute Newsroom 2026-11-15

IBM's new open models learn to ignore orders hidden in documents

IBM's new open models learn to ignore orders hidden in documents

IBM released two small open models on November 12 that mark which text is an instruction and which is data. In IBM's own tests, hidden attacks succeeded less than 6 percent of the time, down from over half. The models are experimental previews, and IBM warns the protection is learned behaviour that should not be the only defence.

Why it mattersAI agents that read email and web pages can be hijacked by text planted there. A built-in separation of commands and content is one of the few structural fixes on offer.
#AI Agents

✓ Verified · 3 sources

WhatsApp X Telegram
Read in the app — free, in 9 languages

Related stories

Kuaishou's new Kling image model generates pictures in native 8K.
2026-11-19
Meta staff fear its Muse agent mixed up users' bank details.
2026-11-19
Arizona opens a criminal probe into OpenAI over rogue bank agents.
2026-11-19
OpenAI says agents finish 46% of its researchers' half-day tasks.
2026-11-19
Google's mid-tier Gemini 4 Neon edges past its own flagship in tests.
2026-11-19