Research
How to Prevent Prompt Injection (13 Attacks, Tested)
First reported · Discovered openrouter.ai
Page published
Earliest dated coverage: 29 Oct 2026 · First observed: 1 Nov 2026 · Latest dated coverage: 29 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
OpenRouter's tests illustrate why tool-enabled agents need externally enforced action checks rather than relying solely on instructions that label documents as untrusted.
OpenRouter reports a controlled prompt-injection study against a document assistant with an email tool across four models. Armor's per-tool-call checks reportedly reduced attacker-induced emails from 73 of 1,440 attempts to zero, while labeling documents as untrusted data reduced successful attacks but did not eliminate them.