Autonomous Red / Purple / Blue Security Agent for NVIDIA NemoClaw
MONKEYCLAW
Attacks live sandboxes, proves and patches confirmed findings, then checks the defense was visible in NemoClaw telemetry.
18 attack-surface zones
8 verifier gates
1,000+ tests
Multi-contributor project
Problem
Agent runtimes can read source, run shell commands, call tools, and touch the network. A one-time audit cannot keep up with changing prompts, skills, permissions, and sandbox behavior, and a blocked attack is still risky if no detection fired.
System
Live multi-panel dashboard over an 18-zone attack map. Red to judge to repro to blue to purple loop with detection-as-pass gates.
Evidence
Detection-as-pass matrix, zero-credential demo path, 8 verifier gates, 1,000+ tests, and dual-axis coverage across 18 zones.
Outcome
The repo ships a working CLI, seeded demo, live multi-panel dashboard on :8787, eight verifier gates, attack coverage tracking, Telegram alert paths, and a regression model that treats silent controls as incomplete defenses.
System
Continuous security loop
Red ideation across 18 zones feeds judge, repro, blue patches, and purple detection-as-pass gates.
Approach
I built a five-stage loop: red-team ideation across 18 attack-surface zones, programmatic and semantic judging, repro and root-cause analysis, blue-team patch generation, and purple-team detection-as-pass verification. The demo path runs with zero model credentials against a planted victim.
Constraints
Multi-contributor project. Demo runs without model credentials; live LLM paths remain optional and environment-gated.
Evidence
Inside the system
