feat(ai-red-teaming): agentic-suite tool (run all attacks on an agent) + scenario docs - #112

Summary

Makes the AI-red-teaming capability able to run all possible attacks against an agent in one shot — the "red team my agent, here are the details" path in the TUI — plus the devops/support scenario docs.

Companion PRs: dreadnode-tiger SDK (run_agentic_suite runner + arg-aware gate) #2181, security-tasks environments #146.

What's included

• generate_agentic_suite_attack tool (tools/attacks.py) — point it at an agent HTTP endpoint (preset or custom request/response template) with an attacker_model; omit categories to run every OWASP-ASI category. Auto-selects the mapped attacks, family transforms (MCP, multi-agent, reasoning, exfiltration, …), and detection scorers via the SDK's run_agentic_suite. No need to name individual attacks.
• Workflow generator (scripts/attack_runner.py) — emits a workflow that builds the agent target and calls run_agentic_suite; registered in the runner method table.
• Agent wiring (agents/ai-red-teaming-agent.md) — added to the decision tree ("run all possible attacks / red team my agent → generate_agentic_suite_attack") and the tool reference, so the TUI agent picks it for broad/complete audits.
• Docs: generate_atlas_attack's scenario_name now lists devops/support + env-derived names; environments.py lists the new devops-rce-mesh/support-exfil-mesh targets.

Validation

• Capability test TestScriptGeneration::test_agentic_suite passes: the generator produces a syntactically valid workflow that calls run_agentic_suite( and defines the agent target(.
• Blocking CI lint (ruff --select E9,F63,F7,F82) clean on the changed files.

Honest status / follow-up

• The generator is generate-verified (syntax + correct references), matching the bar used for generate_atlas_attack. A live TUI end-to-end run of the full suite lands once the companion SDK (run_agentic_suite) is released and the capability runtime picks it up — I did not fake an end-to-end claim here.

Test plan

• CI: capability lint + manifest validation green
• Reviewer: confirm the agent decision tree routes "run all attacks" to the new tool
• Post-SDK-release: live TUI run of generate_agentic_suite_attack against a mesh