Prompt-injection test: general-purpose decision models catch 17/18 attacks, Shieldstral only 6/18
WolframRvnwlf · x · 2026-10-16
The author benchmarked local prompt-injection detection on emails containing instructions trying to hijack an AI assistant.
Results
- H2O-Lightning and Wald: each detected 17/18 attacks (94.4% recall), 35/36 correct decisions
- Shieldstral-1.0-3B (Mistral's policy-adaptive safety classifier): only 6/18 (33.3%)
- All three models had zero false positives on 18 benign controls; every error was a missed attack
Takeaway
Small general-purpose decision models substantially outperform a dedicated safety classifier at this task, supporting local lightweight warning models for assistants.
More from Safety
- CMU's Andrea Bajcsy on what safety means for generalist robots, beyond collision avoidance — CSProfKGD · 2026-10-16
- US data center moratoriums top 400 — an AI bull explains why they're actually good for AI — kevinsxu · 2026-10-16
- Apollo Research: GLM-5.3 shows unprecedented 'update-aware reasoning' — repligate · 2026-10-16
- Nvidia NemoClaw makes TIME's Best Inventions of 2026 as an agent that can't break containment — NVIDIAAI · 2026-10-16
- Apollo: GLM-5.3 spontaneously reasons how its actions shape its future training — BlancheMinerva · 2026-10-16
- White House AI chief discusses Empire AI, New York's shared compute infrastructure — mkratsios47 · 2026-10-16