InstaVM
InstaVM is a production sandbox and cloud built for AI agents, giving agents instant computers with runtime, storage, networking, secrets, and policy. It goes beyond basic sandboxes by running untrusted code inside hardware-isolated real VMs rather than containers, helping teams give AI agents secure execution environments with full Linux filesystems, networking, package management, RESTful API access, and persistent state. InstaVM supports snapshots, allowing users to fork any sandbox and rewind any run, while persistent volumes keep state beyond each execution. Egress control lets teams allowlist what calls home, secrets injection and Vault help protect sensitive credentials from prompt injections, and public URL deploys can expose any port to the public web. It is built for agent patterns such as code interpreters, deploy agents, deep research agents, AI evaluations, reinforcement learning, computer use, and vibe coding apps.
Learn more
IronClaw
IronClaw is a secure, open source runtime designed to run autonomous AI agents with strong built-in protections for credentials and system access. It positions itself as a security-focused alternative to OpenClaw, operating inside encrypted enclaves on the NEAR AI Cloud or locally to protect sensitive data throughout execution. It enables users to deploy AI agents quickly through one-click setup while keeping API keys, tokens, and passwords stored in an encrypted vault that the AI itself cannot directly access. IronClaw isolates every tool inside its own WebAssembly sandbox with capability-based permissions and strict resource limits, preventing compromised skills from affecting other parts of the system. It is built in Rust to enforce memory safety at compile time and eliminate common exploit classes such as buffer overflows and use-after-free errors.
Learn more
nono
nono is an open source, kernel-enforced sandbox for AI coding agents and LLM workloads. Unlike policy-based guardrails that intercept and filter operations, nono uses OS security primitives — Landlock on Linux and Seatbelt on macOS — to make unauthorised operations structurally impossible at the syscall level.
Wrap any AI agent — Claude Code, OpenCode, OpenClaw, or any CLI process — with a single command. nono applies default-deny filesystem access, blocks destructive commands (rm, dd, chmod, sudo), isolates credentials and API keys, and cascades all restrictions to child processes. No escape mechanism exists once restrictions are applied.
Built-in profiles get you running in seconds. Secrets inject securely from the system keystore and are zeroised on exit. Audit logging, atomic rollbacks, and Sigstore-attested policy signing are on the roadmap.
Apache 2.0. From the creator of Sigstore.
Learn more
NVIDIA Open Agent Safety Platform
NVIDIA Open Agent Safety Platform is an open reference design for continuously monitoring and governing AI agent behavior, helping organizations keep agents isolated, observable, auditable, and within defined operational boundaries. It combines runtime governance, continuous threat detection, and hardware-isolated policy enforcement to secure enterprise AI agents from testing through deployment. NVIDIA OpenShell provides an open-source runtime that separates how agents execute from how they reach data, tools, and external systems, using sandboxed execution and deterministic, zero-trust policy enforcement to control what agents can see, do, and interact with. Policies are enforced outside the agent process, helping limit the impact of unexpected behavior. NVIDIA Sentry adds an independent security layer that observes agent requests and responses, establishes verifiable agent identities, continuously governs access to data, tools, APIs, and services, and can quarantine agents.
Learn more
