My test suite was green. My detector was wrong.
I built a scanner to catch poisoned MCP tool descriptions, pointed it at 13,474 real ones, and watched it flag a tool called prompt_injection_scan as prompt injection.
Aug 10, 20269 min read30
Command Palette
Search for a command to run...
Latest articles
Isolate by dependency, not by load
Standard load-testing phases by traffic volume. For an agent with third-party upstreams and multi-call turns, attribution is the harder problem.
Aug 9, 20267 min read8
You can't load-test an LLM agent with a dumb mock
Mocking the model lets you measure everything else. The model was the only thing that mattered.
Jun 4, 20265 min read21