The Open Secure AI Alliance published an August 4 request for comments on its Shared AI Findings Exchange, or SAFE, proposal. The draft would confidentially collect AI security incidents and near misses, notify affected organizations, preserve evidence, and turn recurring control failures into shared defensive guidance; it remains an open proposal rather than an adopted standard.
The Open Secure AI Alliance published a request for comments on August 4 for the Shared AI Findings Exchange (SAFE), a proposed working group for learning from AI security incidents and near misses. The Linux Foundation announcement and the Alliance's draft RFC describe a confidential reporting process intended to notify affected parties and convert recurring failures into evidence-based controls.
SAFE is still a draft for community review. It is not an adopted standard, certification program, or enforcement body.
What the draft asks organizations to report
The RFC proposes five guiding principles: openness with accountability, learning rather than blame, risk-based response, member sovereignty, and separation between confidential learning and legal enforcement. Its reporting compact would cover events such as unauthorized access, escapes from a sandbox or policy boundary, exposure of third-party confidential data, and continued activity after an operator knows a system is outside its approved scope.
The proposed notification schedule is deliberately specific. Directly affected organizations would be notified as soon as possible, customers with credible exposure within 72 hours, and SAFE would receive an initial confidential report within four business days. The draft also proposes a preliminary factual report within 30 days and remediation status within 90 days, subject to security, legal, and investigative constraints.
Evidence across the whole agent stack
The proposal treats an agent as more than a model. Reviews would examine instructions, safeguards, tools, runtime environments, monitoring, human operations, and supply-chain dependencies. Members would preserve prompts, traces, tool calls, logs, configurations, model and safeguard versions, permissions, human approval events, affected artifacts, and a complete incident timeline.
NVIDIA presented SAFE alongside other open agent-security projects at Black Hat, while TechRepublic described the proposal as part of a broader open defense stack. Those contributions provide useful context, but the RFC itself is the authoritative record of what SAFE currently proposes.
What practitioners can use now
Even before the governance model is settled, the draft offers a concrete incident-readiness checklist: define reportable events, preserve reproducible evidence, document human interventions, and set notification deadlines before an agent failure occurs. Whether SAFE becomes operationally important will depend on participation, neutral governance, and organizations' willingness to share enough evidence for others to verify the resulting controls.
Key Points #
- 1SAFE proposes confidential reporting of AI security incidents and near misses, with prompt notification and shared evidence-based controls.
- 2The draft specifies proposed notification milestones and evidence preservation across models, tools, safeguards, environments, monitoring, and human operations.
- 3SAFE remains an open request for comments; adoption, neutral governance, and verifiable participation will determine its practical value.
Scoring Rationale #
The SAFE proposal addresses a real operational gap in reporting and learning from AI-agent security incidents. Its proposed timelines and evidence requirements are relevant to security and ML platform teams, although it remains voluntary draft guidance rather than an adopted standard.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.