Definition
Plain language
A test set for checking whether AI agents can be tricked by hidden instructions planted in their tools or content.
As stated in the literature
A prompt-injection benchmark for LLM tool-use agents, used alongside AgentDojo as a standard measure of single-step injection robustness against current frontier models.
Why it matters: It gives a standard way to measure how easily today's agents can be hijacked by hidden instructions, so defenses can be compared fairly.
For example, it checks whether an agent reading a tool's output will obey a malicious instruction someone slipped into that output.
Heard on the show
“So they take two of the standard prompt-injection benchmarks — AgentDojo and InjecAgent — and they run them against current frontier models with no defense at all.”Episode 105 — The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks