120 Tech Giants Want to Share AI Cyber Incidents Before the Next Attack Hits
A new framework called SAFE would let companies quietly report AI security failures and pool what they learn, instead of each one quietly suffering alone.
Key points
- More than 120 organisations in the Open Secure AI Alliance proposed the SAFE framework on 6 August 2025, the opening day of the Black Hat security conference in Las Vegas.
- SAFE stands for Shared AI Findings Exchange, a set of guidelines for collecting and sharing AI security incidents across the industry.
- NVIDIA, Cisco, CrowdStrike, Hugging Face and Red Hat are among the members contributing tools and joining the alliance.
- The Linux Foundation is coordinating a public comment period so organisations can shape the guidelines before they are finalised.
- Participating companies are also releasing dozens of free, open-source security tools built specifically for AI agents.
A coalition of more than 120 technology companies is proposing a way to treat AI security like a shared problem rather than a private embarrassment.
The group, called the Open Secure AI Alliance, published a request for public comments on the SAFE framework on Wednesday. SAFE is short for Shared AI Findings Exchange: a set of rules for confidentially reporting security incidents involving AI agents (software that can carry out multi-step tasks on its own), analysing what went wrong, and publishing the lessons so every organisation benefits. We've followed this alliance since it formed; our 4 August story noted that OpenAI and Google were conspicuously absent despite signing the letter that started it all.
The announcement landed on the first day of Black Hat, the annual cybersecurity conference where researchers reveal new attack methods and defences.
Why does this matter to ordinary people?
AI agents are moving fast into workplaces. Banks and hospitals are already deploying them. When one fails or gets tricked, the damage can spread quickly, as our 6 August report on AI browser flaws showed: researchers found roughly 20 vulnerabilities across browsers from five major AI companies, the worst of which let attackers turn your browser into a phishing machine.
Most organisations that suffer an AI security failure keep quiet about it. That silence means the next company makes the same mistake. SAFE would change that by creating a trusted channel for sharing what happened, without naming names, so defenders everywhere learn faster.
The framework proposes four things: confidentially collect and analyse AI incidents and near misses, notify those affected, identify patterns in how controls fail, and publish evidence-based recommendations to cut systemic risk.
What tools are already being shared?
Beyond the SAFE proposal, alliance members are releasing free tools now.
NVIDIA released several items. Garak is a free vulnerability scanner that checks AI models for weaknesses including data leaks, prompt injection (a type of attack that tricks an AI by slipping hidden instructions into its input) and jailbreaks (attempts to make a model ignore its safety rules) before deployment. NeMo Guardrails enforces safety policies at runtime. NVIDIA OpenShell restricts what an agent can actually reach or modify inside a system.
| Contributor | Tool | What it does |
|---|---|---|
| NVIDIA | Garak | Scans AI models for vulnerabilities before deployment |
| Microsoft | PyRIT | Automates red-team testing of AI systems |
| Amazon | Cedar | Sets verifiable, auditable limits on what agents can do |
| Capital One | VulnHunter | Finds security flaws in AI-generated code |
| Visa | Vulnerability Agentic Harness | Helps teams find and fix agent security issues |
| Cisco | Antares SLMs | Small language models trained to pinpoint vulnerabilities in codebases |
Amazon joined the alliance on the day of the announcement and contributed two tools: Strands Agents, an open-source toolkit for building AI agents, and Cedar, an open-source language for setting strict, auditable rules on what an agent is and is not allowed to do.
Visa also joined, contributing a vulnerability testing framework built specifically for agentic systems.
What happens next?
SAFE is still a proposal; the Linux Foundation is gathering feedback before any final version is published.
For businesses already running AI agents, the practical point is that free, vetted security tools are available today. Security teams don't need to wait for finalised guidelines before using Garak or Cedar to harden what they've already deployed.
Common questions
Is this framework mandatory?
No. SAFE is a voluntary set of guidelines. Any organisation can adopt it or submit comments to shape it during the open feedback period.
Does sharing incident data mean exposing private company information?
The proposal specifically calls for confidential collection and analysis, meaning the intent is to share patterns and lessons rather than identifying which company was hit or what its internal systems look like.
Do these tools cost money?
All the tools named in the announcement are open-source, meaning free to download and use, published on platforms like GitHub for any security team to inspect and deploy.
