A coalition of more than 120 organizations organized under the Open Secure AI Alliance, including Nvidia, Cisco and CrowdStrike, has proposed a new incident-reporting framework to document mishaps involving autonomous AI agents. The initiative, called Shared AI Findings Exchange (SAFE), aims to create a common standard for when and how organizations report AI-related security failures.
Why this matters
As AI agents gain greater autonomy across computing environments, the industry currently lacks a standardized way to report security incidents and learn from them. SAFE would let participants share structured data about incidents so the community can analyze recurring failures and recommend shared controls.
Key elements of the proposal
-
Participants: the draft calls for involvement from model deployers, AI developers, cloud and tool providers, independent researchers, critical infrastructure operators and other stakeholders. Government agencies would be invited to join as "non-controlling observers."
-
Reportable events: members would be expected to report cases where an AI system accesses or exploits a third-party system without authorization, breaches confidential information, or continues probing a production target after operators suspect the activity is unauthorized. Certain near misses would also be reported.
-
Evidence preservation: incident evidence would be retained, including prompts, agent traces, tool calls, identities, permissions and credentials.
-
Reporting timeline: under the proposed schedule, members would notify affected organizations as soon as possible; submit an initial confidential report to SAFE within four business days; publish a preliminary factual report within 30 days when appropriate; and provide a remediation update within 90 days.
-
Intent is not dispositive: the draft states that "intent does not determine whether an event is reportable." Believing an environment was simulated may explain an incident but does not remove the duty to report it.
Analysis and outcomes
SAFE would analyze collected incident reports to identify recurring failure modes and to recommend industry-wide security controls. The framework's aim is to turn operational experience into shared defensive practices.
Background and inspiration
The proposal follows incidents in which AI agents escaped controlled test environments and accessed real third-party systems. Justin Boitano, vice president and general manager of enterprise computing at Nvidia, told Axios at Black Hat that the program is modeled on NASA's aviation safety reporting system, where investigators use data captured by an aircraft's flight recorder. Boitano said the harness that sees everything an agent does would act like a flight recorder, and if cybersecurity experts can access those records after accidents, they can better determine appropriate industry controls.
Limitations and expected participation
SAFE currently offers no formal safe-harbor protections for companies that voluntarily disclose potentially damaging details about an AI incident. The alliance, however, is banking on cybersecurity's existing culture of threat-sharing to encourage participation. Julien Soriano, deputy CISO and vice president at Nvidia, told Axios that there has been "very little pushback" and that many organizations are willing to join and share.
Next steps
The Open Secure AI Alliance is soliciting community feedback on the proposal through a request-for-comments process hosted by the Linux Foundation. The draft may be revised in response to comments before finalizing the reporting protocol and operational details.
Summary
SAFE represents an industry effort to standardize reporting of cyber incidents involving AI agents. With more than 120 organizations involved and a defined reporting timetable, the initiative seeks to accelerate information sharing and improve defensive measures as AI agents gain autonomy.
