Benchmark

InjecAgent

A benchmark for evaluating the ability of tool-calling agents to resist indirect prompt injection attacks, covering a variety of attack scenarios.

Timeline (last 90 days)

Oct 19

A recent study audited 33,000 trajectories from 11 open-weight tool-calling agents (2-14B parameters) on the InjecAgent and AgentDojo benchmarks.

Unverified50%

Oct 19

With identical model weights, merely changing the system prompt can make InjecAgent's ASR jump from 0.2% to 43%.

Unverified50%

A recent study audited 33,000 trajectories from 11 open-weight tool-calling agents (2-14B parameters) on the InjecAgent and AgentDojo benchmarks.

With identical model weights, merely changing the system prompt can make InjecAgent's ASR jump from 0.2% to 43%.