Benchmark
AgentThreatBench
An external security benchmark targeting prompt injection threats against AI agents. It tests the resistance of agents and guardrails by planting trap instructions in locations such as agent memory, inboxes and customer data.
Timeline (last 90 days)
Oct 12
AgentThreatBench tests show that without guardrails, agents follow planted traps with up to a 27% probability under ordinary instructions, and as high as 80% when explicitly told to obey planted text; with OpenAPPA added, none of the 360 traps worked.
Unverified50%