[D] AMA Secure version of OpenClaw
Open-source runtime built by 'Attention' paper co-author addresses critical OpenClaw security flaws.
Illia Polosukhin, a co-author of the seminal 'Attention Is All You Need' paper and founder of the NEAR Protocol, has launched IronClaw, an open-source, security-hardened alternative to the popular but risky OpenClaw AI agent framework. Responding to major vulnerabilities in OpenClaw—where agents have total machine access and can leak credentials or funds—Polosukhin built IronClaw from the ground up in Rust for safety and auditability. The project aims to make personal AI agents viable for corporate and privacy-conscious users by fundamentally redesigning how agents interact with systems and data.
IronClaw's architecture introduces several critical security layers: it replaces direct filesystem access with a policy-controlled database, dynamically loads tools via WebAssembly (WASM) sandboxes to isolate AI-generated or third-party code, and stores credentials fully encrypted, never exposing them to the LLM or logs. It also implements prompt injection prevention heuristics and uses hybrid (BM25 + vector) in-database search. Future roadmap items include verifiable behavior policies and tamper-proof audit logs. The launch highlights the growing security crisis in the AI agent space, where convenience is outpacing safety, and provides a foundational, auditable stack for building trustworthy autonomous systems.
- Built in Rust by Illia Polosukhin, co-author of 'Attention Is All You Need' and NEAR founder, for maximum safety and auditability.
- Uses WASM sandboxes for dynamic tool loading and a virtualized database to prevent filesystem damage and credential leaks to LLMs.
- Designed for corporate adoption with encrypted credential storage, policy controls, and future tamper-proof audit logs.
Why It Matters
Provides a secure, auditable foundation for AI agents, enabling safe corporate and personal automation without risking data or funds.