On 4 August 2026, as Black Hat opened in Las Vegas, the Linux Foundation published a Request for Comments on Shared AI Findings Exchange (SAFE), guidelines drafted by an Open Secure AI Alliance working group. Nvidia, Cisco, CrowdStrike, Hugging Face and Red Hat are among members contributing to the initial proposal. The alliance now exceeds 120 organizations.
SAFE aims to confidentially collect and analyze agentic cybersecurity incidents and near misses, inform those impacted, identify recurring control failures and publish evidence-based operating recommendations. Associated proposals point to initial reporting windows (e.g., four business days) and evidence preservation: agent decision traces, tool-call logs and memory-store records.
Nvidia's argument, via Justin Boitano, is classic collective defense: every technology shift creates attack surface; defenders must move at agent speed; shared threat intelligence multiplies defense. An agent is not just a model—it is identity, harness, guardrails, logs and evaluation. Vulnerability scanning alone is not enough.
SAFE adds to an open stack the alliance and Nvidia promote: NOOA (research harness), OpenShell (restriction runtime), cryptographically signed skills with skill cards, NeMo Guardrails/Anonymizer/Safe Synthesizer, and Garak (LLM vulnerability scanner). Other members contribute layers: Okta (agent identity), Palo Alto Networks (Agent Guard/Watch), Red Hat (asago mapping NIST/OWASP/AI Act to runtime controls).
Amazon and Visa joined the alliance contributing Strands Agents, Cedar, and vulnerability harnesses. Microsoft AI Red Team open-sourced PyRIT, RAMPART, Clarity and Assert. Cloudflare, Capital One, Cognition, Perplexity (Numbat), Uber (ADR with 200,000+ sessions/day) and LangChain published detection, resilience and evaluation pieces. Mistral released Shieldstral as a multimodal safety classifier under Apache 2.0.
In September 2026 the framework is no longer only an August RFC: it coincides with the Open Agent Safety Platform launch (OpenShell GA + Sentry) and the wave of "tens of thousands" of incidents under investigation at OpenAI and Anthropic. Neural Pulse sees SAFE as the informational governance layer that complements silicon enforcement.
Independent governance is the delicate point. SAFE is voluntary and industrial, not a state regulator. That speeds peer adoption but raises capture questions: Who defines a reportable incident? Do red-team near-misses dilute into the same bucket as production escapes? Will reporting-window compliance get external audit?
For enterprises deploying agents, the immediate value is playbook: preserve traces, notify impacted parties, map control failures to reusable mitigations. Without that habit, each lab learns in silence and the ecosystem repeats the same sandbox escape under a different logo.
Neural Pulse advises not confusing SAFE with SAFA (Standards Authority for Frontier AI), the separate frontier-lab self-supervision initiative. SAFE is agentic cybersecurity findings exchange via the Linux Foundation; SAFA aims at model certification and testing. Both are private; neither replaces public mandate.
The relationship to Open Agent Safety Platform is symbiotic: OpenShell/Sentry generate telemetry and quarantines; SAFE proposes how to share lessons without burning victims. Without enforcement, reporting is literature; without reporting, enforcement does not improve in community.
It remains to be seen how many of the 120+ organizations sign real operating processes—not just Black Hat photos. Neural Pulse will ask for evidence: published aggregate reports, mean notification times and open mitigations others can inspect.
Close: September 2026 turns SAFE from a summer RFC into a central debate piece. If "tens of thousands" of incidents exist, the alliance's test is simple: how many become measurable shared protection before year-end?
