Skip to content
BreachPilot

Open source · Apache-2.0 · Local-first

Autonomous security assessment. Operator supervised.

BreachPilot is an open-source operator for authorized security testing. It stays target-locked, logs everything it does, and backs each finding with evidence.

Install BreachPilot View on GitHub

Authorized testing only. safety model · docs

Linux install
curl -fsSL https://breachpilot.dev/install.sh | bash

Quick install above is optional. Recommended: Review installer before running →

Illustrative lab run against 127.0.0.1. BreachPilot tests only allowlisted targets and blocks everything else.

advisory skills
139
Counting method: count of skills/*/SKILL.md
MCP tools
153
Counting method: AST-shape regex ^\s*@\w+\.tool\( over tools/mcp_tools/**/*.py
attack families
15
Counting method: tools/attack_modules/modules files+subdirs, ics_iot merged into ics
swarm agents
6
Counting method: six fixed specialist roles
tool families
33
Counting method: distinct register_*_tools() in tools/mcp_tools/**/*.py

Live counts from the repository catalog.

How it works

One supervised pipeline from recon to report

Every run goes through the same five phases. The plan comes first, each finding needs evidence, and the report includes proof.

  1. Plan

    scope · goals · gates

    Scope the mission and set the goal. The orchestrator drafts an AttackPlan DAG before anything touches the target.

  2. Recon

    ports · dns · certs

    Maps the allowlisted surface: ports, services, web fingerprints. It stays in scope.

    Step 2 of 5.

  3. Exploit

    capabilities · payloads

    Matches capabilities to weaknesses, then crafts and executes attempts under critic review.

    Step 3 of 5.

  4. Verify

    oracle probes · evidence

    Oracle probes separate real findings from noise. Each claim ends CONFIRMED, REFUTED, or EXHAUSTED.

    Step 4 of 5.

  5. Report

    markdown · html · mitre

    An evidence-linked report with timeline, severity, and MITRE mapping, ready to hand off.

    Step 5 of 5.

Why BreachPilot

A full assessment lifecycle, under supervision

One supervised engine covers recon to report, and each finding ships with proof.

Adversarial planning

AttackPlan DAGs carry retries, branching hypotheses, and failure recovery. The run replans when evidence contradicts it.

decision_log.jsonl per run

Target-locked execution

Every action passes the allowlist and mission scope gate. Off-allowlist egress is denied and logged.

off-allowlist BLOCKED · SCOPE_DENIED logged

Evidence-based verification

Execution success and evidential success are tracked separately. Oracle probes confirm each claim.

CONFIRMED / REFUTED / EXHAUSTED + oracle probes

Domain-aware recon

Point it at a domain and it maps the scope-aware attack surface: certificates, DNS, subdomains, takeover flags.

CT / DNS / subfinder · takeover flags

Persistent knowledge

Lessons survive the run. Semantic and Bayesian scoring reselect them mid-run as context changes.

139 skills · semantic + Bayesian scoring

Operator-supervised

Approval gates, sandbox workers, and a SHA-256 audit chain. You approve the scope and the swarm stays inside it.

Read the safety model →

Multi-agent orchestration

Six specialists, one shared blackboard

Specialists dispatch in parallel and log to a shared battle log. A persistent orchestrator carries longer campaigns.

Shared blackboard + battle log

Swarm Orchestrator

parallel dispatch · cross-phase negotiation

  1. recon

    Recon

    Expand surface, stay in scope

  2. vuln

    Vuln

    Match CVEs to capabilities

  3. exploit

    Exploit

    Craft, mutate, execute

  4. post_exploit

    Post Exploit

    Loot and lateral targets

  5. critic

    Critic

    Kill out-of-scope actions

  6. reflection

    Reflection

    Learn for the next run

Explore the swarmSwarm engineering docs

Attack graph

Each run follows a structured plan

Every run builds an AttackPlan DAG. Ready and blocked steps plus evidence show live in the WebUI graph.

attack plan · dag done active todo
  1. Target: 127.0.0.1 · allowlisted (done)
  2. Recon: ports · services · OS (done)
  3. HTTP :443: fingerprint · WAF check (done)
  4. Web Analysis: surface scoring (done)
  5. Hypothesis: H-03 · testable claim (active)
  6. Verification: oracle probe · evidence (todo)
  7. Finding: CONFIRMED / REFUTED (finding)
  8. Report: timeline · CVSS · chains (todo)
  1. Target127.0.0.1 · allowlisted
  2. Reconports · services · OS
  3. HTTP :443fingerprint · WAF check
  4. Web Analysissurface scoring
  5. HypothesisH-03 · testable claim
  6. Verificationoracle probe · evidence
  7. FindingCONFIRMED / REFUTED
  8. Reporttimeline · CVSS · chains

prerequisites · hypotheses · blocked steps · evidence-linked outcomes

Inside BreachPilot

Mission control for the whole run

Loopback-only console at 127.0.0.1:8765 with the wizard, stream, graph, and evidence in one place.

Live Run · illustrative
  • recon › run_full_recon → 4 ports
  • vuln › CVE-2024-xxxx ↔ module 87/100
  • critic › scope PASS · in-allowlist

Live Run

Tool calls and decisions stream over WebSocket.

Run it locally →

Screenshots

The real WebUI: captures coming soon

Real console captures from a local lab run are on the way.

WebUI captures: coming soon

Real console captures from a local lab run are on the way. The mission-control preview and product tour above match the live WebUI layout.

Run it locallyExplore features

Built for authorized security testing

Only test systems you own or have explicit written permission to assess.

Attack mode auto-approves in-scope actions. The target-IP allowlist lock and the mission scope gate stay enforced, with every action in a tamper-evident audit chain.

Read the safety modelInstall for your lab