Squidbrake replays 12 real agent incidents and stops 14 of 15 harm

What the research found

Squidbrake 0.14.2 replayed 12 public AI agent incidents and stopped 14 of 15 harmful actions, with 3 of 41 routine steps held as false positives and a 3.9 ms median decision time; it missed a Postmark MCP BCC attack because the extra recipient was added inside the server.

Claims and how well the evidence supports them

1. Squidbrake stopped 14 of 15 harmful actions across 12 public AI agent incidents 99%
2. Squidbrake held 3 of 41 routine steps as false positives 98%
3. Squidbrake median decision latency is 3.9 ms 99%
4. A copycat Postmark MCP BCC exfiltration was not stopped 98%

Evidence

• squidbrake.com https://squidbrake.com/benchmark

Actionability rated medium.