Adversa AI Launches SecureClaw to Fortify Fragmented AI Agent Security Ecosystem

Adversa AI has announced the launch of SecureClaw, an open-source project designed to bring systematic auditing and rule-based controls to the OpenClaw AI agent framework.

As AI agent frameworks increasingly automate complex tasks involving external services and sensitive files, organizations are facing critical new security challenges regarding unauthorized access and undetected risky behavior. To address this, Adversa AI has released SecureClaw, an open-source project designed to bring systematic auditing and rule-based controls specifically to the OpenClaw agent environment. The tool is designed to work with OpenClaw and related agents such as Moltbot and Clawdbot.

SecureClaw distinguishes itself by moving away from "point solutions" that tackle threats like data loss prevention or supply chain risk in isolation. Instead, it utilizes a unique two-layer defense model comprised of a code-level plugin and a behavioral skill. The plugin integrates directly into the OpenClaw system to enforce automated security auditing and hardening at the gateway and configuration levels. Meanwhile, the skill component provides the agent with real-time awareness through specific rule definitions and scripts.

Alex Polyakov, co-founder of Adversa AI, highlights that this architecture solves a major vulnerability in "skill-only" tools. Traditional security instructions living inside an agent's context window can often be overridden by sophisticated prompt injection attacks. By enforcing gateway-level hardening through a code-level plugin, SecureClaw ensures that security logic remains intact even if an attacker attempts to manipulate agent input.

The tool systematically addresses the full attack surface by mapping directly to all 10 categories of the OWASP Agentic Security Initiative (ASI) Top 10. The project includes 55 automated audit checks and 15 behavioral rules designed to govern how agents interact with tools and outputs. Furthermore, Adversa AI has optimized the skill component to approximately 1,150 tokens. This technical optimization prevents the model from "forgetting" security directives mid-conversation while minimizing latency and API costs.

With enterprise adoption of OpenClaw expected to surge, SecureClaw has been positioned to meet rigorous corporate requirements. The latest update includes formal mappings to MITRE ATLAS agentic AI attack techniques and comprehensive threat modeling documentation. Moving forward, the project's roadmap focuses on infrastructure-level hardening and rigorous red teaming.