We put 10 models through the same attacker harness, against two agents with 23 known vulnerabilities. π½
The cheapest model found 16 of them in three runs.
The most expensive found about 17 in one run, for 15 times the cost.
Link in comments.
We're open to every agent builder for the next two weeks. πΎ
"Test my agent for free" is live. Point Enoki at your agent's endpoint and it starts attacking: inputs, memory, tool calls, retrieval, outputs.
Free, no install, no credit card.
Link in comments.