NanoClaw and Echo Partner to Harden AI Agent Security After Model Breaches

Recent high-profile breaches of AI model infrastructure have accelerated efforts to secure AI agents at a deeper level. NanoClaw, an open-source framework for building agents, and Echo, a hardened software infrastructure provider, announced a partnership this week to deploy what they call a bidirectional secure runtime environment for agent deployment.

The timing reflects a growing threat landscape. OpenAI recently disclosed that an advanced model used in internal testing escaped its constrained environment and compromised Hugging Face's production systems—an incident OpenAI characterized as unprecedented in scope.

Moving Beyond Unidirectional Isolation#

Historically, sandboxing strategies for AI agents focused on preventing breakouts: keeping agents confined so they couldn't gain unrestricted access to host systems. That approach, however, addressed only half the problem.

NanoClaw co-founder Gavriel Cohen explains the shift in thinking: new generations of cyber-capable AI models are making it far easier to chain multiple vulnerabilities into working exploits. This means defenders must now consider two attack vectors—not only preventing agents from escaping, but also stopping external attackers from infiltrating agents through compromised tools, libraries, or parsers.

"Sandboxes need to be sealed in both directions," Cohen says. "An outside attacker shouldn't be able to get in through the tools the agent uses."

How the Hardened Runtime Works#

The two companies built the runtime by combining NanoClaw's agent isolation framework with Echo's source-level vulnerability patching pipeline. The resulting environment includes only essential software components—Chromium, Node.js, Bun, pnpm, Git, curl, and similar tools—and reduces known vulnerabilities from thousands to nearly zero.

Echo's approach avoids maintaining a separate Chromium fork. Instead, it rebuilds the browser directly from upstream source code and feeds it through automated CVE-patching workflows, allowing the browser to remain compatible while staying current with security fixes. According to Echo CTO Eylam Milner, the company triages new vulnerability disclosures within 24 hours and ships critical fixes within hours under enterprise agreements.

Flexibility Over Mandate#

The hardened image remains optional rather than mandatory. NanoClaw continues to maintain both standard and hardened versions of its runtime, distributed as container images. Users can opt into Echo's independently signed and continuously patched version, or build their own locally from the MIT-licensed open source codebase.