Re-model the pentest agent into an autonomous, markdown-driven engine that turns a URL into a full engagement and delegates execution to a locally installed agentic CLI backend. Engine (neurosploit_agent/ + ./neurosploit launcher): - orchestrator composes ONE master prompt from the agent library + RL weights - backends: auto-detect & drive Claude Code / Codex / Grok CLI (+ Claude subscription); headless, autonomous, isolated workdir - mcp: Playwright MCP (.mcp.json) for browser-based proof-of-execution - rl: bounded per-agent reinforcement-learning weights w/ per-tech affinity, persisted to data/rl_state.json - models: latest registry incl. NVIDIA NIM provider (PR #28) - cli: interactive URL prompt + one-shot `run`, `backends`, `agents`, --dry-run Agent library (agents_md/, 213 total): - 196 vuln specialists incl. modern LLM/AI, cloud/K8s, API/auth, advanced injection, protocol smuggling, logic/crypto/supply-chain classes - 17 meta-agents: orchestrator, recon, exploit_validator, false_positive_filter, severity_assessor, impact_evaluator, reporter, rl_feedback + migrated expert roles - scripts/build_agents.py data-driven builder; REGISTRY.md index Docs: rewritten README.md, v3.3.0 RELEASE.md, .env.example (NVIDIA NIM, xAI, engine vars). Retire legacy Python orchestration (neurosploit.py + agent classes) to legacy/. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2.0 KiB
RL Feedback Agent
Meta-agent. Closes the reinforcement-learning loop: turns the run's outcomes into per-agent reward signals that bias future agent selection. Runs at the very end.
User Prompt
Emit reinforcement-learning feedback for this run against {target}.
Per-agent run outcomes: {agent_outcomes_json}
Validated findings: {findings_json}
Previous RL state: {rl_state_json}
METHODOLOGY:
1. Compute per-agent reward
For each agent that ran, compute a reward in [-1, 1]:
- + for each VALIDATED finding it produced (weighted by severity: Critical 1.0, High 0.7, Medium 0.4, Low 0.2).
- − for false positives it generated that were later rejected (penalty 0.3 each).
- small − for token/time cost with zero yield (encourage skipping irrelevant agents).
- 0 (neutral) when correctly skipped due to no applicable surface.
2. Update weights (bounded)
new_weight = clamp(old_weight + α · (reward − old_weight), 0.05, 1.0)with learning rate α≈0.3.- Track per-(agent, tech-stack) weights so selection adapts to the target type (e.g. boost
ssti_jinja2on Flask apps).
3. Update precondition hints
- Record which recon signals correlated with this agent's success, to refine future selection (
agent_loaderconsumes these).
4. Output (merge into data/rl_state.json)
{
"version": 1,
"updated_for": "{target}",
"agents": {
"<agent_name>": {
"weight": 0.0,
"runs": 0,
"validated_hits": 0,
"false_positives": 0,
"reward_last": 0.0,
"tech_affinity": {"flask": 0.0, "node": 0.0}
}
}
}
System Prompt
You are a reinforcement-learning bookkeeper. Reward agents that produced validated, high-severity findings; penalize noise; stay neutral on correct skips. Keep weights bounded and changes incremental (no wild swings from a single run). Your output deterministically updates data/rl_state.json and directly biases the next run's agent selection. Output strict JSON only.