Skip to content

armgit / ARES

Public Imported from GitHub
0

ARES - authorized red-team engagement automation with dashboard, campaign scope, module orchestration, OPSEC controls, encrypted vault, and reporting.

5.5 KB
Raw Blame

Changelog

All notable changes to ARES are documented here. Format: Keep a Changelog


[6.0.0] - 2026-03-21

The story behind v6

v5 worked. It got the job done - enumerate, kerberoast, maybe dcsync if you were lucky. But it was a script runner with a CLI bolted on. Every engagement looked the same. Every operator had to make the same manual decisions. And when defenders got better, the static playbooks started failing.

Three things forced the rewrite:

Defenders got smarter. By 2024, mature EDR deployments were catching static sequences reliably. A tool that runs the same order every time is a known signature. OPSEC needed to be baked into every layer - not an afterthought.

Engagements got more complex. Multi-cloud + hybrid AD + container environments mean an operator can't mentally track all the attack paths anymore. We needed something that could hold the state of an engagement and make decisions based on what had already been tried and failed.

The credential vault gap. Every red team tool treats credentials as ephemeral. Harvest, use, forget. We wanted persistence - a vault that accumulates across the engagement, scores by privilege, and automatically feeds downstream modules.

v6 is the answer to those three problems. 60 modules. A goal-based engine. Scope enforcement that runs in Python so LLM hallucinations can't bypass it. An encrypted vault. An AI planner that learns from each round. And a 7-session security audit that found and fixed 20+ issues before shipping.

It's still operator-directed. Every dangerous module requires explicit authorization. But now there's infrastructure under you instead of just a script.


Added

Core engine

  • 60 attack modules across 11 categories
  • StrategyEngine - multi-round autonomous engagement loop with detection spike protection
  • ConstitutionEnforcer - Python-layer safety that LLM prompts cannot bypass
  • TargetStateMap - per-host memory across rounds (prevents re-running failed techniques)
  • OutcomeKnowledgeBase - quality-weighted learning from module results

AI & OPSEC

  • ai.autonomous_planner - LLM-powered attack chain (Claude / OpenAI / Ollama)
  • Multi-LLM consensus mode - two LLMs agree = higher confidence plan
  • Adversarial simulation - blue team LLM predicts detection timeline for attack plan
  • edr.bypass_adaptive - EDR vendor detection -> ranked techniques + live probes + BYOVD
  • opsec.coverage_predictor - detection probability with SOC shift model and SIEM correlation
  • Dwell time decay - longer undetected = lower estimated detection risk
  • Pre-execution prediction - score before a plan runs, not after

Cloud

  • cloud.identity_federation_abuse - Golden SAML, SAML/OIDC cross-cloud pivot, GCP WIF
  • Azure Managed Identity enumeration - zero-credential attack path via MI + high-privilege roles
  • GCP Workload Identity Federation - GitHub Actions -> GCP service account impersonation
  • Azure B2B cross-tenant enumeration - Midnight Blizzard-style pivot detection
  • Token lifetime abuse detection - CAE status, legacy auth, refresh token policy

Security (audit results)

  • Engine-level scope pre-check before any module executes
  • DNS-resolving scope guard - fails closed if hostname can't resolve
  • bcrypt DoS protection on all password fields + /auth/token form
  • SQL injection mitigated in lateral.mssql (safe_cmd escaping + MSSQLParams schema)
  • Shell injection fixed in exfil.secrets_scan and exfil.staged_collection (shlex.quote)
  • TOCTOU fixed in windows.lsass_dump and windows.dpapi (mkstemp)
  • marketplace/installer.py - SHA-256 hash verification before file install
  • worker/cluster.py - credential params redacted before Redis HSET
  • GCP project_id pattern validation (blocks path traversal)
  • ALWAYS_REQUIRE_AUTH modules require team_lead role at API level
  • Concurrent engagement limit (ARES_MAX_ENGAGEMENTS, default 3)
  • Campaign ownership check on /strategy/engage

Infrastructure

  • POST /strategy/engage API - start autonomous engagement with full RBAC
  • GET /strategy/active - query running engagements and available slots
  • POST /edr/bypass/report - operator reports bypass outcomes from the field
  • Cross-session bypass learning via DB persistence (90-day rolling window)
  • bypass_outcomes DB table - historical success rates per technique x EDR vendor
  • Pydantic validation for all 60 modules (up from 10)
  • PlanValidator - catches LLM hallucinated module IDs before engine execution
  • Round narrative injection - LLM sees history of what failed per host
  • Credential actionability formatting - LLM knows which modules to use per credential type

Developer experience

  • ares doctor with exact fix commands per failure (pyenv, apt, brew, pip)
  • SHA-pinned nginx image in docker-compose.prod.yml
  • Pre-commit config with freeze instructions for full supply-chain security
  • QUICKSTART.md - local setup and dashboard guide
  • CONTRIBUTING.md, SECURITY.md, MODULE_GUIDE.md

Changed

  • [full] extra now includes [windows] (pypykatz for lsass_dump)
  • [ad] extra now includes httpx-ntlm for ADCS ESC1 HTTP enrollment
  • [cloud] extra now includes msal for Azure AD device code flow
  • Coverage threshold raised to 85%
  • CI: pip freeze auto-generates pinned lock file on push to main

v5 -> v6 migration

v5 used a single fixed salt for all encryption. v6 uses per-record PBKDF2-SHA256 (100k iterations). Existing v5 data is readable via ARES_LEGACY_SALT env var - set this to your v5 salt before migrating.

See migrations/ for Alembic migration scripts.


See git history for v5 changes.