Ironclaw is an open-source runtime designed to execute artificial intelligence agents inside secure enclaves on NEAR AI Cloud. Written in Rust for compile-time memory safety, the system provides isolated environments to prevent prompt injection attacks from stealing sensitive user credentials. It manages API keys and passwords using an encrypted vault that injects secrets only at the host boundary, ensuring that underlying large language models never directly view raw credentials. Additionally, Ironclaw executes external tools within WebAssembly sandboxes enforcing capability-based permissions, which mitigates the risk of malicious third-party code exfiltrating data or accessing unauthorized files. The runtime inspects outgoing data using real-time outbound traffic scanning and restricts destinations through network allowlisting. Ironclaw also leverages hardware-based Trusted Execution Environments to ensure agent data stays encrypted in memory during execution. The platform is model-agnostic, enabling integration with major language model providers. Ironclaw is intended for artificial intelligence developers, DevOps engineers, privacy-conscious enterprises, and security researchers seeking private, verifiable execution environments. It is distributed as an open-source project and is freely available for use and deployment.
Problem: AI agents often require API keys or passwords to perform tasks, but these secrets can be exposed to the model through prompt injection attacks.
Solution: Ironclaw uses an encrypted vault that injects credentials only at the host boundary. This ensures the raw secrets are never visible to the AI model itself.
Example: An agent managing a user's GitHub and Slack accounts without the LLM ever having access to the raw OAuth tokens.
Problem: Using community-developed AI skills or plugins carries the risk of malicious code exfiltrating data or accessing unauthorized files.
Solution: Every tool in Ironclaw runs within its own WebAssembly (Wasm) sandbox, isolated from the rest of the system with strict capability-based permissions.
Example: Running a data-scraping tool from a public repository while ensuring it cannot access local files or make unauthorized network calls.
Problem: Organizations want to deploy AI agents in the cloud but are concerned about the cloud provider or external actors accessing sensitive data in memory.
Solution: Ironclaw deploys agents within Trusted Execution Environments (TEEs) on the NEAR AI Cloud, where data remains encrypted even during processing.
Example: A healthcare assistant processing patient records in a secure enclave that prevents the cloud host from viewing the data.
Target audience: Best for: AI Developers, DevOps Engineers, Privacy-conscious Enterprises, Security Researchers
Pricing: Open Source
Ironclaw is an open-source security-focused runtime that executes artificial intelligence agents in encrypted enclaves on NEAR AI Cloud. Built with Rust for memory safety, it keeps sensitive data protected using encrypted secret vaults, WebAssembly tool sandboxing, real-time outbound traffic scanning, and network allowlisting. It works with major large language model providers.
Ironclaw is an open-source tool and is free to use. Users can download, modify, and deploy the runtime according to its open-source licensing terms without paying licensing fees to access its core memory-safe runtime, sandboxing, and secure enclave integration features.
To install and set up Ironclaw, visit the project official website and repository documentation at https://ironclaw.com/. The documentation outlines prerequisites, environment requirements for NEAR AI Cloud enclaves, and step-by-step instructions for compiling and configuring the Rust-based runtime.
Ironclaw uses an encrypted vault that injects credentials exclusively at the host boundary rather than exposing them directly inside model prompts. This architectural separation prevents language models and potential prompt injection exploits from accessing raw passwords, API keys, or OAuth tokens while still enabling the agent to authenticate with external services.
Ironclaw executes external plugins and community tools inside dedicated WebAssembly sandboxes. Each sandbox enforces strict capability-based permissions, which blocks unapproved local file access and unauthorized network calls. Combined with outbound traffic scanning and network allowlists, this sandboxing approach prevents malicious code from exfiltrating sensitive organizational information.