The problem
Millions of people are deploying AI agents with zero security posture. Unvetted skills execute on every heartbeat. Credentials sit exposed. Prompt injection works on every default configuration.
Nobody is building the security layer. We are.
What We're Building
These tools are in design phase — grounded in our published research, not yet shipping code. Sign up below to get notified when they launch.
AgentArmor Coming Q2 2026
Runtime behavioral monitoring for AI agents. Will detect anomalous tool calls, permission escalation, and data exfiltration in real time.
SkillVet Coming Q2 2026
Designed to scan third-party agent skills for malicious patterns and permission violations. Behavioral sandbox — will run the skill, not just scan it.
RedClaw Coming Q2 2026
Automated red-team toolkit for AI agents. Will let you point it at your agent and get a vulnerability report with OWASP Agentic mapping.
Research
Our tools are built on published research, not assumptions.
Agent Red-Team Framework
7 attack classes against AI agents. 100% reasoning chain hijack rate on default configurations. 19 scenarios across prompt injection, tool manipulation, and memory poisoning.
Read the findings →RL Agent Vulnerability
162 attacks across 40 trained RL agents. Observation perturbation 20-50x more effective than reward poisoning. Mapped to OWASP Agentic Top 10.
Read the findings →govML
ML governance framework. 50+ templates, 10 profiles, 20+ generators. Contract-driven reproducibility across 9 projects with 469+ tests.
Methodology →Get notified when we launch.
Open-source agent security tools. Early access and input on what we build next.
