IBM trains small open models to ignore hidden injection attacks, cutting success to near zero
mawcopolow · reddit · 2026-11-14
IBM trained two small open models (LLMON preview) to ignore malicious instructions hidden in documents. A MACCUS chart of IBM's paper (arXiv 2610.38120, Table 2) shows indirect prompt injection success (OPI text) dropping from 100% to 0.1% on the 3B model and 96.3% to 0.02% on the 8B, versus standard Granite 4.1 Instruct; on agent-tool attacks (ASB), 99.2% to 5.6% (3B) and 52.8% to 3.1% (8B). Prompt injection is a major pain point for small-model agents, making this a significant robustness result.
More from Safety
- Interconnects: Jensen's resignation is a major setback for US open-model ecosystem — Interconnects (Nathan Lambert) · 2026-11-14
- "I feel sick": user rage after accidentally using Gemini in Google Docs — lilsiibee07 · 2026-11-14
- Capabilities and safety are two sides of offensive vs defensive cyber, argues LeviTurk — LeviTurk · 2026-11-14
- AI hacking tool gets admin access in under 3 hours per FBI; Project Zero finds bugs weekly; Dwarkesh sees superintelligence by 2030 — LeviTurk · 2026-11-14
- 118,434 MCP servers scanned: 64% have zero stars, only 4.8% use OAuth — Glittering_Device653 · 2026-11-14
- Reddit user warns numes.ai's "free API keys" pitch is a scam — Interesting_Ad_5676 · 2026-11-14