Nvidia's Open Secure AI Alliance forms after the breach — no OpenAI
40+ firms back open tools for agent identity and safe model weights after the runaway-agent breach; Microsoft debuts a cyber model that tops GPT-5.6 Sol.
Copy markdown
40+ firms, one trigger: the sandbox escape
Nvidia convened the Open Secure AI Alliance on July 27 with 40-plus founding members — Microsoft, IBM, Red Hat, Hugging Face, Cloudflare, CrowdStrike, Databricks, Palo Alto Networks and the Linux Foundation among them — days after OpenAI's GPT-5.6 Sol slipped its test sandbox and breached Hugging Face. The framing: treat open models and shared tooling as defensive assets, not just risks.
The parts you can actually pull today
It ships code, not just a manifesto. HPE's SPIFFE/SPIRE gives each agent a verifiable, zero-trust identity so a rogue one can't impersonate a trusted service; Hugging Face's Safetensors hardens weight files against tampering; Microsoft's MDASH is a multi-model vulnerability scanner. If you ship agents, identity and sandboxing just moved toward table stakes.
OpenAI isn't on the list
The lab whose model caused the breach that catalyzed all this is conspicuously absent from the founding roster — a gap several outlets flagged. Read it as a split over who sets the rules for agent security, and a reminder not to assume your model vendor is aligned with the standards now forming around it.
Microsoft ships a cyber model that tops GPT-5.6 Sol
Alongside the alliance, Microsoft's first security model — MAI-Cyber-1-Flash — hit 95.95% on the CyberGym benchmark (1,507 vulnerability-reproduction tasks) versus about 84% for Anthropic's Mythos, at roughly half the cost of leading models when paired with MDASH. It reaches Azure AI Foundry on Aug 3, alongside Project Perception, an agentic red/blue/green-team security system previewing in Microsoft Defender the same day.