Building a safer AI future, together

Security is the worst possible place to create AI lock-in.

During the recent Hugging Face incident, closed AI systems blocked essential forensic work because they could not distinguish defenders from attackers.

Hugging Face needed to investigate and contain the attack immediately. It could not wait for a provider to change its safeguards. So the team ran an open-weight frontier model on its own infrastructure, analyzed more than 17,000 actions, and contained the intrusion.

That is the practical case for open models in cybersecurity.

The lesson is not that open models are inherently safer than closed ones. Defenders need both. But when an organization’s systems are under attack, it needs the freedom to choose the right model, inspect and adapt it, protect sensitive data, and run it under its own control.

AI security also depends on far more than model weights. It requires secure harnesses, identity, permissions, guardrails, logs, and evaluation across the entire agent stack. Making more of that stack open allows a global community of defenders to test it, strengthen it, and respond faster.

Faros supports the mission of the Open Secure AI Alliance. Attackers will use the most capable AI available. Defenders must have access to equally capable tools—including open systems they can control when closed ones cannot help.

We congratulate Nvidia, Hugging Face, The Linux Foundation, Microsoft, CrowdStrike, Cloudflare, and the many organizations contributing to this effort.