TL;DR
SAFE is a draft plan from the Open Secure AI Alliance and the Linux Foundation. It asks member companies to report AI security incidents and near misses on a set schedule, keep early details private, then publish lessons so others can defend themselves. Right now it is only a Request for Comments. There is no live system to report to yet. The idea is strong, but the value depends on who actually joins and shares.
Introduction
When a plane has a close call, the airline files a report. Other airlines read it. Everyone gets safer. Cybersecurity has never worked that way. When an AI agent goes off the rails at one company, the next company learns nothing until it happens to them too.
A group of more than 120 companies wants to change that.
On August 4, 2026, the Linux Foundation published a draft called the Shared AI Findings Exchange, or SAFE. It is a proposed set of rules for sharing what goes wrong with AI agents, so the whole industry can learn from one bad day instead of repeating it.
What SAFE Actually Is
SAFE stands for Shared AI Findings Exchange. A working group inside the Open Secure AI Alliance wrote it. The Linux Foundation, a neutral group, runs the review.
The goal is simple to say and hard to do. Take a real AI security incident. Study it in private. Warn the people affected. Spot the failures that keep repeating. Then publish guidance others can use.
Members Nvidia, Cisco, CrowdStrike, Hugging Face, and Red Hat helped write the first draft.
One thing to be clear about. SAFE is a Request for Comments. That means it is a draft open for public feedback on GitHub. There is no timeline for launch. There is no way to file a report yet. It is a starting point, not a finished rule.
The Reporting Clock
The draft asks members to move fast when something goes wrong. According to reporting from Cybersecurity Dive on the RFC, the proposed timeframes look like this:
- Notify customers of a credible data exposure within 72 hours
- Report the incident to the exchange within 4 business days
- Publish a preliminary report within 30 days, subject to legal and security limits
These numbers come from the draft and could change during the comment period. Treat them as a proposal, not a final policy.
Why Share Incidents At All
The model comes from older industries. Aviation and finance have shared incident data for decades. A mistake at one airline becomes a safety rule for all of them.
AI agents make this kind of sharing urgent. An agent is not just a model. It has identity controls, tools, logs, and the power to act across your systems. When one fails, the failure can look nothing like a normal software bug.
Right now, when an agent misbehaves at one company, that lesson stays locked inside that company. The next target starts from zero. SAFE tries to break that pattern by turning one incident into shared defense.
The Catch Nobody Should Skip
An incident exchange is only as good as what its members share.
Reporting from various sources note that OpenAI, Google, and Anthropic are not listed as Open Secure AI Alliance members. Those are three of the largest AI labs. If the companies most likely to generate incident data sit outside the room, the shared pool of lessons has real gaps.
That is not a reason to dismiss SAFE. It is a reason to watch who signs up as the draft moves forward. A framework like this earns its value from breadth. The more voices inside it, the more useful its warnings become.
What This Means for Your Security Team
You cannot report to SAFE today. But you can get ready for a world where AI incident sharing is normal.
Start by asking a plain question. If one of your AI agents did something it should not, would you even know? Could you reconstruct what it touched, in what order, and why?
That is the real work SAFE points to. Not the paperwork, but the ability to see what your agents actually do. Teams that already log and trace agent behavior will find any future reporting rule easy. Teams that do not will scramble.
