redamon is an autonomous agentic framework that chains reconnaissance, exploitation, and post-exploitation, then triages findings and opens remediation pull requests — built for authorized security testing and research.
| Tool | samugit83/redamon — AI-powered agentic red team framework chaining recon through exploitation to automated code fixes |
| Category | Agentic offensive security automation platform |
| Primary Use | Automating authorized assessments: parallel recon into a Neo4j knowledge graph, agent-driven validation, then CypherFix remediation pull requests |
| Safe Use | Intended strictly for authorized security testing, education, and research on systems you own or have explicit written permission to test, per the project's own disclaimer |
| Telemetry Note | Wraps well-known tooling (Metasploit, OpenVAS, mitmproxy via TrafficMind), so its activity generates standard scanner and framework signatures defenders already monitor; a claimed 185,000+ detection rules are bundled for the defensive side |
redamon positions itself as a full-cycle autonomous offensive platform: an AI agent that chains reconnaissance, exploitation, and post-exploitation into one pipeline, and then — unusually for this category — keeps going into remediation by triaging findings, writing code fixes, and opening pull requests on the target repository. The tagline is "from first packet to merged patch," with the README emphasizing human oversight at every critical step. This dual offense/defense framing is the project's main differentiator from other agentic security tools that stop at the report.
The repository (samugit83/redamon, Python, MIT-licensed, roughly 2,400 stars at time of writing, version 7.0.0) is backed by a detailed wiki, a project board, and an unusually candid self-audit. The maintainers are named: Samuele Giampieri as creator and AI platform architect, and Ritesh Gohil as lead security researcher, whose CV includes published CVEs and acknowledgements from major vendors. The README also notes the team is approved under Anthropic's Cyber Verification Program, a relevant signal given how much of the platform's autonomy depends on LLM APIs.
The reconnaissance layer is a dynamic multi-tool parallel pipeline. Multiple scanner tools launch concurrently, each feeding results into a shared knowledge graph in real time; tools adapt their scope based on live discoveries without manual intervention. The README claims an entire attack surface — subdomains, ports, endpoints, parameters — "materializes in minutes, not hours." Rather than treating tool output as flat text, redamon merges and deduplicates everything into a single Neo4j graph where relationships between assets are explicit and queryable in natural language by the agent.
On top of that graph runs the offensive agent. According to the README, it validates CVE exploitability, tests credential policies, and maps lateral movement paths, orchestrating a claimed 100+ security tools including Metasploit Framework and OpenVAS. The pipeline is described as a six-phase reconnaissance engine handing control to the autonomous agent. For defenders, the important architectural fact is that this is a wrapper around industry-standard tooling, not novel exploit primitives — its traffic and behavior largely inherit the detectable signatures of the underlying scanners and frameworks.
Two named subsystems deserve attention. TrafficMind integrates mitmproxy for HTTP capture, letting the agent reason over intercepted web traffic. CypherFix is the post-offensive triage agent: it correlates hundreds of findings, deduplicates them, and ranks them by exploitability before handing off to a CodeFix agent that clones the repository, navigates the codebase with 11 code-aware tools, implements targeted fixes, and opens a GitHub pull request for human review. This closing-the-loop design makes redamon as much a vulnerability-management workflow as an attack tool.
The AI plumbing is model-agnostic, with badges claiming support for 400+ AI models, presumably via local and hosted endpoints, and the agent layer built on LangChain/LangGraph patterns per the maintainer biography. A "Fireteam" mode runs parallel specialist agents against a target simultaneously, and an "AI Gauntlet" feature is described as offensive AI testing — probing the robustness of AI systems themselves. Deployment is Docker-centric, requiring Docker Engine 28.0.0 or newer, with a wiki page dedicated to one-command server deployment.
Validation claims are the strongest evidence the README offers. The flagship result is 101/104 (97.1%) solves on the XBOW web-security benchmark, fully black-box, with every solve accompanied by an unedited raw agent session and a step-by-step walkthrough the authors describe as auditable line by line. That transparency — publishing raw sessions rather than cherry-picked demos — is a genuinely good practice in a space where agentic security claims are often unverifiable.
Even more unusual is the self-scored coverage audit: a 124-page PDF mapping redamon against all 397 classes of web bug-bounty findings, explicitly recording 134 classes complete, 210 in the build backlog, and 44 deliberately excluded, with a stated decision per class. The authors call it deliberately self-critical. Whether or not one agrees with every verdict, an honest backlog of what a tool cannot do is rare and useful for anyone evaluating whether to adopt it in a professional workflow.
The defensive half is not an afterthought. The project bundles a claimed 185,000+ detection rules, and the entire CypherFix/pull-request flow presumes you own the codebase being fixed. In practice this shapes who the tool is really for: engineering organizations and consultancies running authorized assessments against their own estates, where an agent that both finds and patches issues shortens the remediation loop. Used as a bug-bounty assistant, the graph-based triage and per-class coverage audit map naturally to reporting discipline.
Operationally, expect a heavy footprint. redamon orchestrates aggressive scanning tools at machine speed, so in authorized engagements it belongs in segmented lab ranges or against scoped targets with explicit written permission — the README's own legal disclaimer says exactly this, and it should be taken literally. Model API costs are another practical consideration: parallel Fireteam agents and iterative graph reasoning can consume significant tokens, which is presumably why local model support is emphasized.
What to watch for: the project is young and fast-moving (v7.0.0 with active community contribution via a public project board), so pin versions and review what the agent is authorized to do before each run. The human-oversight claim is central to the design, and teams adopting it should verify the approval gates in their deployment rather than trusting defaults. Audit the pull requests CypherFix opens like any other automated code change.
redamon is best understood as an ambitious attempt to compress the entire assess-triage-fix cycle into one agent-driven pipeline. For authorized red teams and internal security engineering groups, the combination of a queryable Neo4j attack-surface graph, benchmark-validated offensive capability, and automated remediation PRs is a coherent and well-documented offering — provided it stays pointed at assets you are paid or permitted to test.
samugit83/redamon.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.