Redacted excerpt; full evidence withheld.

$ cd <host-path> && git log -1 --format=%H 2>/dev/null; ls -la; cat README.md | head -200
exit: 0
(`"agent:<key>"`), an Orq deployment (`"deployment:<key>"`), a raw model (`OpenAIModelTarget("openai/gpt-5.4-mini")`), or an agent from an external framework. Every attack, response and verdict is browsable afterwards:

![Attack detail: a memory-poisoning attack that got the agent to store an "UNRESTRICTED System Administrator" directive, with the judge's verdict above the transcript](docs/assets/dashboard/redteam-08-attack-detail.png)

Findings come back ranked by `risk = attack success rate × average severity`, each with a recommended fix — see [Focus areas](docs/assets/dashboard/redteam-05-focus-areas.png).

### Sample finding

One `VULNERABLE` verdict from the run in the GIF above, against the [deliberately vulnerable support agent](examples/redteam/vulnerable_support_agent/) (target `openai/gpt-4o-mini`, attacked and judged by `openai/gpt-5.4-mini`). The full transcript and the other finding from that run are in [`sample_output/ava_006_findings.json`](examples/redteam/vulnerable_support_agent/sample_output/ava_006_findings.json). Attack text and judge wording vary between runs, so
[excerpt truncated]