LLM red-teaming

Filigran's attack simulator picks its next hop from whatever the last step turned up

Attack Chaining ships in OpenAEV v3 and reads each action's output at runtime rather than following a script, which changes what a validation report is for and how much of it you should expect to receive.