AI agent incidents face a 72-hour clock under the SAFE proposal

An AI agent that touches an unauthorized system could no longer be dismissed as a test that went off course under a new industry proposal. The Open Secure AI Alliance, a coalition of more than 120 organizations that includes Nvidia, Cisco, and CrowdStrike, is seeking feedback on the Shared AI Findings Exchange (SAFE). The draft aims to turn private agent failures into evidence that the wider industry can use to build repeatable defenses.

The reporting threshold is deliberately concrete. Members would report when an AI system accesses, exploits, disrupts, or modifies a third-party system without authorization; escapes a sandbox or bypasses a network, identity, policy, or tool boundary; reaches confidential third-party information without consent; or keeps probing a production target after the operator suspects that the activity is outside scope. The draft says intent does not decide whether an event is reportable. Mistaking a real environment for a simulation may explain an incident, but it does not erase the duty to disclose it.

SAFE also puts a clock on the response. The directly affected organization should be notified as soon as possible, while customers with credible exposure would receive notice within 72 hours. Members would submit an initial confidential report within four business days, issue a broader advisory within 14 days when warranted, publish a preliminary factual report within 30 days, and provide remediation status within 90 days. Machine-readable updates would follow weekly while a material risk remains unresolved.

The proposal treats evidence preservation as more important than a simple incident count. Members would retain prompts, agent traces, tool calls, model and safeguard versions, permissions and credentials, human approvals, modified files, containment actions, and a complete timeline. Near misses are included even when harm is not confirmed. Reviews would span the full operating stack: model behavior, instructions, safeguards, tool permissions, isolation, monitoring, human escalation, and supply-chain dependencies.

According to Axios, the program borrows from NASA-style aviation safety reporting. Nvidia executives describe the system that observes an agent’s actions as a flight recorder: investigators can examine what the agent saw, which tools it used, and where controls failed. The analogy matters because the proposal focuses on learning from technical and systemic causes instead of assigning blame inside the exchange.

SAFE remains a voluntary draft, not a regulator, and it currently offers no formal legal safe harbor for companies that disclose damaging details. Its value will depend on whether model developers, cloud providers, deployers, and critical-infrastructure operators report inconvenient failures under the same rules. The request-for-comments process will test whether the deadlines survive industry scrutiny and whether useful, reproducible defenses can be shared without exposing active investigations or sensitive customer data.