Leading technology companies have joined an international effort to improve how the industry reports and responds to security incidents involving autonomous artificial-intelligence agents.

On 4 August 2026, the Linux Foundation published a Request for Comments on the proposed Shared AI Findings Exchange guidelines, commonly known as SAFE.

The proposal is being developed through the Open Secure AI Alliance, a group that now includes more than 120 organisations. NVIDIA, Cisco, CrowdStrike, Hugging Face and Red Hat are among the members contributing to its initial development.

Amazon and Visa were also announced as new contributors to the alliance’s expanding collection of open AI-security tools.

What the SAFE guidelines propose

The SAFE initiative is intended to establish a trusted process through which organisations can share information about AI security incidents without unnecessarily exposing confidential or commercially sensitive details.

Its proposals include mechanisms to:

  • Collect and analyse AI-agent security incidents confidentially
  • Report near misses before they become major breaches
  • Notify organisations and individuals affected by an incident
  • Identify security-control failures that appear repeatedly
  • Publish evidence-based operating recommendations
  • Turn lessons from one incident into protection for other organisations

This process could give developers and security teams access to information that helps them recognise emerging attack patterns before the same weaknesses are exploited elsewhere.

Why AI agents introduce new risks

AI agents differ from traditional chatbots because they can act on a user’s behalf. Depending on their permissions, they may access business applications, retrieve files, modify databases, execute code or communicate with external services.

An agent is therefore more than an AI model. Its security also depends on its identity controls, tools, runtime environment, permissions, activity logs and operating guardrails.

If any of these components is incorrectly configured, an agent could access restricted information or perform an action beyond its intended responsibility.

The alliance argues that scanning an AI model for vulnerabilities is insufficient. Companies must also monitor how the complete agent system behaves when connected to real tools and information.

Technology companies contribute defensive tools

Members of the alliance are contributing several open-source tools covering different parts of the AI-security system.

NVIDIA highlighted the following technologies:

  • NVIDIA OpenShell, which restricts what information and systems an AI agent can access.
  • NOOA, a research harness intended to make agent behaviour easier to test, trace and audit.
  • Garak, an open-source scanner for data leaks, prompt-injection weaknesses and jailbreak attacks.
  • NeMo Guardrails, which helps developers apply safety policies to AI applications.
  • Verified agent skills that are scanned, documented and cryptographically signed.

Cisco has contributed technologies including DefenseClaw, its agent-governance layer, and Antares security language models. Microsoft has open-sourced tools such as PyRIT, RAMPART and Assert for AI risk identification, red-team testing and safety evaluations.

Other contributors include Cloudflare, Okta, Palo Alto Networks, Capital One, CrowdStrike, Uber, Visa, Veeam and LangChain.

Proposal remains open for public comments

SAFE is currently a proposed framework, not a final or compulsory international standard. The Linux Foundation’s Request for Comments allows cybersecurity professionals, AI developers, companies and other interested parties to examine the recommendations and submit feedback.

The consultation process will help determine how incidents should be classified, what information can be shared and how confidentiality should be protected.

If widely adopted, the guidelines could become an important foundation for collective cyber defence as autonomous AI systems gain greater access to corporate networks and critical digital infrastructure.