Project Feral
An open security-research initiative threat-modeling real-world agentic AI — applying OWASP ASI, CSA MAESTRO, and MITRE ATLAS to systems that are already in the wild.
Threat-modeling agentic AI, in the open.
Project Feral is SecuraAI's independent analysis of OpenClaw, an open-source AI-agent platform that passed 200,000 GitHub stars in two months. OpenClaw grants agents autonomous access to shell commands, file systems, messaging, and 100+ integrations through the Model Context Protocol — and its explosive growth outpaced its security posture, with thousands of instances exposed publicly within days.
It is one of the first practical applications of the OWASP Agentic Security Initiative (ASI) Top 10 (2026), the CSA MAESTRO 7-layer architecture framework, and the MITRE ATLAS taxonomy to a real-world agentic system — and the entire body of work is published openly for educators and researchers.
An architecture-level threat model.
The model held up against real incidents.
After release, the Phase I threat model was cross-referenced against real-world incidents and OpenClaw patches. The findings were borne out.
- CVE-2026-25253 · CVSS 8.8
One-click RCE via token exfiltration — maps directly to OC-T01 (prompt injection) and OC-T02 (unsandboxed execution).
- ClawHavoc campaign
335 malicious skills found in the ClawHub marketplace — validating OC-T05 (supply chain), which was upgraded from High to Critical.
- MITRE ATLAS · Feb 2026
Four new case studies (AML.CS0048–0051) and seven new agentic-AI techniques added to the ATLAS framework.
A tri-framework approach.
Three complementary standards, so findings speak the language security teams already use.
OWASP ASI Top 10 (2026)
The risk taxonomy for agentic-AI applications — the language security teams use to classify what can go wrong.
CSA MAESTRO
A 7-layer architectural decomposition for multi-agent systems, used to locate each threat in the stack.
MITRE ATLAS
Adversarial tactics, techniques, and procedures for AI/ML systems — tying findings to documented real-world cases.
Read the research.
Architecture Threat Model
The complete architecture-level threat model — 10 threats, 5 attack chains, 6 trust boundaries, with full OWASP ASI and CSA MAESTRO mapping and a prioritized remediation roadmap.
Read the report ↗Phase I.5Real-World Validation
A delta analysis cross-referencing the threat model against real incidents and OpenClaw patches — 7 of 10 threats validated, with MITRE ATLAS case-study matches.
Read the report ↗Free for educators & researchers.
All Phase I and Phase I.5 materials are published under CC BY-NC-SA 4.0 for non-commercial educational use — built for course case studies, student projects, research baselines, and framework validation.
License & disclosure ↗Test it with us.
SecuraAI is recruiting academic collaborators, security researchers, and industry partners for hands-on testing against an isolated OpenClaw environment — red teaming, vulnerability scanning, defensive tooling, and comparative analysis.
Register for Phase II ↗Explore the full research portal.
Threat browser, attack chains, methodology, and the complete reports — open and free.
Architecture-level analysis based on public source code and documentation. Project Feral is not a penetration test, vulnerability scan, or formal security audit, and SecuraAI maintains no affiliation with the OpenClaw platform maintainers. Provided for educational and defensive security purposes.