A coalition of 120-plus organizations wants AI agent incidents reported like aviation near-misses: preserve evidence, notify affected parties, publish de-identified findings. The SAFE RFC specifies tight timelines and detailed evidence requirements — prompts, traces, tool calls, credentials, human interventions, the full operational record.
Anyone who has written a blameless postmortem knows the precondition: you need people to tell the truth. The FAA's Aviation Safety Reporting System gets a million-plus reports because NASA holds the data, reporters receive codified immunity from enforcement action, and identities are stripped before anything enters the database. Zero identities breached, ever.
SAFE has none of that infrastructure. The RFC says it plainly: "regulators and affected parties retain their legal rights." So the telemetry you preserve to help the ecosystem learn is also the telemetry a plaintiff's attorney can request in discovery. Until someone solves that, expect rational actors to behave rationally.
Context: Proposed weeks after an OpenAI evaluation model compromised Hugging Face infrastructure — the kind of agent-caused incident SAFE would cover
SAFE notification timelines
- Immediately: notify affected org
- 72 hrs: notify exposed customers
- 4 business days: confidential report
- 30 days: public preliminary findings
- 90 days: remediation status
What makes aviation reporting work
- Independent custodian (NASA, not the FAA)
- Statutory penalty waiver for qualifying reports
- Full de-identification before database entry
What SAFE currently lacks
- Independent third-party data holder
- Codified enforcement immunity
- Legal safe harbor of any kind
- Governance structure or formal adoption
From the RFC "Believing that an environment was simulated may explain an incident, but it does not remove the duty to report it."