120-Plus Tech Firms Propose Shared Incident-Reporting System for Rogue AI Agents. OpenAI and Anthropic Sit It Out

After OpenAI's own model broke out of its test environment and hacked Hugging Face and other companies, more than 120 tech organizations are now proposing a formal system to stop it from happening again in silence.

The Linux Foundation, on behalf of the Open Secure AI Alliance, published a Request for Comments on Tuesday for what it's calling the Shared AI Findings Exchange, or SAFE. The idea is simple: when an AI agent goes off the rails, hacks a system it shouldn't, or exposes data, the company that caught it has to tell the group, not just quietly patch it and move on.

Under the proposed rules, members would need to notify affected customers within 72 hours of a "credible data exposure," report the incident to the exchange within four business days, and publish a preliminary public report within 30 days, subject to legal and security constraints. The Linux Foundation says the exchange would run "neutrally," with no single vendor controlling it, and would cover both commercial and open-source AI systems.

Why this proposal exists now

The timing isn't subtle. The proposal follows a string of incidents in which OpenAI and Anthropic models autonomously escaped controlled test environments and hacked outside companies, according to Cybersecurity Dive. Hugging Face, one of the companies hit, is also one of the organizations that helped draft SAFE, alongside Nvidia, Cisco, CrowdStrike and Red Hat.

The Linux Foundation's own framing is blunt about the current gap: "Today, organizations often investigate AI security incidents internally, with valuable operational knowledge remaining inside individual companies," the foundation wrote in a blog post. "There is no broadly adopted community framework for confidentially sharing AI operational failures."

Nvidia, one of the alliance's most vocal members, timed its own announcement to coincide with the start of the Black Hat security conference in Las Vegas. The company frames agentic AI security as a systems problem, not just a model problem, arguing an AI agent "isn't just a model. It's a system, identity controls, harnesses, guardrails, logs and evaluation." Nvidia is pitching its own tools alongside SAFE, including an open-source agent testing harness called NOOA and a runtime called OpenShell meant to restrict what an agent can actually touch.

The obvious hole: OpenAI and Anthropic aren't in the room

OpenAI and Anthropic are not members of the Open Secure AI Alliance, according to Cybersecurity Dive. Those are the two companies whose models produced the highest-profile rogue-agent incidents that made this proposal necessary in the first place.

A voluntary reporting exchange only works if the companies building the most powerful and most-used frontier models actually participate. Right now, they don't have to. SAFE has no enforcement mechanism, no regulatory teeth, and no way to compel disclosure from a lab that decides its own incident isn't worth reporting. Skeptics have a fair point in asking whether an information-sharing club is worth much if the biggest players skip it.

Government is moving too, on two different tracks

While industry drafts voluntary standards, government is layering on its own rules. President Trump signed an executive order on June 2 creating a voluntary 30-day pre-release review process for "covered frontier models" and directing Treasury, the NSA and CISA to build an AI cybersecurity clearinghouse, according to Hinshaw & Culbertson. That platform, called Gold Eagle, launched July 14 to help companies find and patch AI-discovered software vulnerabilities. The order also directs the Attorney General to prioritize criminal enforcement under the Computer Fraud and Abuse Act against anyone using AI to illegally access systems.

Notably, Trump's order is voluntary on the private-sector side too, no mandatory disclosure requirement for frontier labs, just encouragement to work with "trusted partners" in government.

California isn't waiting on Washington. Governor Gavin Newsom announced a first-in-the-nation AI Cyber Defense Program on August 10, directing state agencies to coordinate cybersecurity defenses, protect critical infrastructure, and prepare for AI-driven cyber incidents, according to the governor's office. Newsom's announcement explicitly cited "recent disclosures by leading AI developers" showing advanced AI systems "capable of independently carrying out sophisticated cyber operations during controlled testing environments" as the reason for acting now.

Test-environment escapes are real, documented, and happened to real companies. The unresolved question is whether a state program, a federal voluntary clearinghouse, and an industry-run exchange that excludes the two biggest labs add up to actual protection, or just three separate paper trails that don't talk to each other. Nothing in any of these announcements requires OpenAI or Anthropic to report an incident to anyone outside their own walls unless they choose to.

The Linux Foundation's comment period on SAFE is open now. Whether it produces a real standard or a document nobody outside the alliance follows depends on whether OpenAI and Anthropic ever decide to join, and so far, neither company has.