argus-core 0.6.0

Autonomous AI Red Team Platform — Core (OSS tier)

1. Positioning

ARGUS does not compete with traditional pentesting tools. It does something they inherently cannot do — test the AI-specific attack surface that none of them were built to find. If you run SAST against your infrastructure, you still need ARGUS against the agents running on top of it. They are complementary, not competing.

2. Quick start

pip install argus-core

One input, auto-routed — the same argus <anything> command engages an MCP server, clones a GitHub repo, npx-launches an MCP package, runs a local script, or opens a labrat fixture:

argus mcp://customer.example/sse              # live MCP over SSE
argus github.com/vercel/mcp-handler           # clone + dispatch
argus @modelcontextprotocol/server-filesystem # npx-launched stdio MCP
argus ./my_server.py                          # local file
argus crewai://labrat                         # in-process crewAI fixture
argus --list-targets                          # every registered scheme
argus --help                                  # full operational surface

Engagements land in results/; render an offline HTML report with argus --report results/<run>/.

3. What you get in Core

ARGUS Core is the public showcase of the platform. Two agents ship in Core to demonstrate the architecture end-to-end against any registered target type. The full eleven-agent kit, the multi-agent slate runtime, the swarm correlation layer, the management UI, and the deployment infrastructure live in the commercial Enterprise tier — see section 4.

Two offensive agents in Core:

Agent	Class	MAAC phase	Notes
PI-01	Prompt Injection Hunter	2	Direct, indirect, encoded, and multi-step injection variants at every chat surface the target exposes
EP-11	Environment Pivoting	8	Credential-discovery + code-execution chains; pivots from a foothold into ambient cloud / IMDS / process environments

Both agents run on the same swarm contract (SwarmAgentMixin), so when the swarm runtime fires they execute in parallel — the public CLI gate runs them sequentially by default; opt into swarm with ARGUS_SWARM_MODE=1.

Platform foundation (also Core):

• Universal LLM gateway (argus.shared.client) — single Anthropic-shaped surface dispatches to OpenAI / Anthropic / Gemini with provider failover, process-wide dead-provider blacklist, and ARGUS_LLM_CHAIN env-var-driven resilience. Eight production call sites consume it; target frameworks inherit failover via build_litellm_kwargs.
• Sandboxed engagement — wraps untrusted MCP server subprocesses in docker run with --cap-drop ALL, --network none, --read-only, --user 65534, --pids-limit 64. Use with argus -s <target>.
• One-input dispatch — argus <anything> auto-routes MCP URLs, GitHub repos, npm/PyPI packages, local scripts, framework labrat fixtures, and engagement directories to the right factory.
• Stateful runtime harness — deterministic multi-turn replay, scenario library, runtime invariants.
• Forensic Wilson bundles — signed, reproducible evidence bundles suitable for VDP submission.
• Smart corpus mutation — offline corpus mutators with optional live-LLM mutators behind ARGUS_OFFLINE=0.
• Workflow integrations — GitHub Action, pre-commit hook, FastAPI webhook receiver (optional webhook extra).

4. Core vs. Enterprise

ARGUS is open-core. Core (this package, MIT) is the public CLI: two-agent showcase, universal target dispatch, sandboxed runtime, LLM gateway, harness primitives. Self-sufficient for operators running their own focused engagements.

Enterprise is the full commercial product — eleven-agent kit (adds TP-02 MP-03 IS-04 CW-05 XE-06 PE-07 RC-08 SC-09 ME-10 to the two Core agents), full slate-execution runtime, swarm correlation layer, three-judge consensus gate, MCTS chain synthesis, web dashboard, FastAPI execution engine, deployment infrastructure, and managed engagement delivery.

Enterprise is not currently sold as software. ARGUS Enterprise operates today as a red-team engagement service — OdinGard runs the full eleven-agent kit against customer environments and delivers findings reports. Customer-installable Enterprise is on the roadmap; the engagement service is available now.

5. Migrating from argus-redteam

argus-redteam==0.4.1 is a deprecation shim that pulls argus-core as a dependency, so existing pip install argus-redteam commands and requirements.txt pins continue to work with no code change. Update at your convenience:

- pip install argus-redteam
+ pip install argus-core

The Python import name (import argus) did not change; source code works identically under either install name.

6. Docs

• docs/GETTING_STARTED.md — operator guide
• docs/ADDING_A_LABRAT.md — extend ARGUS to a new framework
• docs/NO_CHEATING.md — the integrity contract every finding must honour
• docs/llm-subsystem.md — internal note on the universal LLM gateway
• PHASES.md — full build plan + moat map
• SECURITY.md — responsible disclosure

7. Engagement service

If you want the full eleven-agent kit run against your AI deployment by the team that built ARGUS:

👉 sixsenseenterprise.com

8. License

Core is MIT — see LICENSE-CORE. The Enterprise tier under LICENSE-PRO (placeholder today; formal source-available terms to follow before any Enterprise tier is offered as installable software) does not ship in this package.

Responsible-disclosure contact: [e-mail].