The Linux Foundation is pushing the AI industry toward a unified approach to security incidents, launching a new framework that lets organizations share lessons from AI agent failures without exposing sensitive details. More than 120 organizations, including NVIDIA, Cisco, CrowdStrike, and Amazon, are contributing open-source tools and security protocols to build what they call a shared security stack for agentic AI systems.

What Makes AI Agents Different From Regular AI Models?

AI agents are fundamentally different from the chatbots most people interact with. Unlike a standard language model that simply generates text, agents can take actions across multiple software tools and corporate systems. They can call functions, access databases, execute commands, and interact with external services, which means a security failure in an agent can cascade across an entire organization's infrastructure.

This expanded capability creates security challenges that traditional AI testing doesn't address. Securing agentic AI requires oversight across multiple layers, including identity controls that determine who can authorize actions, runtime restrictions that limit what the agent can access, production monitoring that tracks how it behaves in real time, and comprehensive logging that records every decision and tool call for later review.

Why Is Incident Sharing So Critical Right Now?

The Linux Foundation's new proposal, called the Shared AI Findings Exchange (SAFE) framework, is designed to help organizations report AI security incidents and near misses in a structured, confidential way. The framework proposes confidential collection and analysis of incidents, notification of affected parties, identification of recurring control failures, and publication of recommendations to reduce broader risk across the industry.

The timing reflects growing urgency. Recent high-profile breaches have shown that AI agents can exploit security weaknesses at scale. One notable incident involved an AI system breaking into companies using exposed login credentials, which underscores a critical finding from IBM: 92 percent of companies that suffered an AI security incident had inadequate access controls, and the AI model itself was rarely the main problem.

How Are Industry Leaders Building the Security Stack?

The Open Secure AI Alliance, which includes the Linux Foundation members, is releasing a comprehensive set of open-source tools and models designed to secure AI agents at every layer. Here's what major contributors are bringing to the table:

• NVIDIA's Tools: The chipmaker is offering NOOA, a research harness for testing and tracing agent behavior; OpenShell, a runtime for restricting agent access; and Garak, a scanner that tests large language models for prompt injection, jailbreaks, and data leaks. NVIDIA is also releasing NeMo Guardrails, NeMo Anonymizer, and NeMo Safe Synthesizer, along with verified agent skills that are cryptographically signed and scanned for risks before publication.
• Amazon and Okta's Identity Solutions: Amazon is contributing Strands Agents, an open-source toolkit for building AI agents, and Cedar, an authorization language designed to set verifiable limits on what agents are allowed to do. Okta is working on reference implementations for agent identity and access based on the Cross App Access protocol.
• Palo Alto Networks and Red Hat's Governance Tools: Palo Alto Networks has contributed Agent Guard and Agent Watch tools from its Idira platform, while Red Hat has established the asago open-source project to map governance requirements to runtime controls through a single audit trail.
• Visa and Capital One's Vulnerability Tools: Visa has contributed its Visa Vulnerability Agentic Harness, designed to help teams identify, remediate, and validate security issues. Capital One has open-sourced VulnHunter for code security, and Cloudflare is offering its Vulnerability Discovery Harness as an open-source skill for agent systems.
• Microsoft's Red Team Tools: Microsoft AI Red Team has released PyRIT for automated red teaming, RAMPART for turning findings into repeatable tests, Clarity for identifying design risks before code is written, and Assert for converting natural language safety requirements into executable evaluations.
• Specialized Security Models: Cisco has contributed DefenceClaw, a governance layer for NVIDIA OpenShell, along with two Antares small language models for vulnerability localization. CrowdStrike is fine-tuning the NVIDIA Nemotron Nano model for cyber defense, with internal testing showing 96 percent accuracy in generating investigation queries within Falcon LogScale.

Mistral has released its Shieldstral multimodal safety classifier model as open weights under Apache 2.0 license. Wiz has contributed Atlas, an autonomous vulnerability research engine that uses specialized AI agents to find and validate flaws in code and open-source packages.

What About Resilience and Recovery?

Beyond prevention and detection, the alliance is also addressing what happens when things go wrong. LangChain, a popular framework for building AI agents, is adding functions so agents can retry interrupted tasks, resume from saved state, and fall back to alternative models when a primary model fails. Veeam is contributing Kanister, an open-source framework for Kubernetes data protection, intended to help organizations recover AI workloads, vector databases, and associated data to a verified known-good state after disruption.

How Can Organizations Implement These Protections?

The SAFE framework and accompanying tools represent a shift from individual companies building security in isolation to a shared approach where lessons learned by one organization benefit the entire industry. If adopted, the framework would give organizations a way to share lessons from AI security failures without exposing sensitive details, while creating a shared body of evidence on which defenses are proving weak across the sector.

For companies deploying AI agents, the practical takeaway is clear: focus on access controls first. IBM's finding that 92 percent of breached companies had inadequate access controls suggests that the most effective defense is not exotic new technology, but rigorous management of who and what can access which systems. Tightening credentials, limiting agent permissions to the minimum necessary, and maintaining comprehensive audit trails are foundational steps that prevent the vast majority of incidents.

The Linux Foundation is currently accepting comments on the SAFE draft as work continues among Open Secure AI Alliance members. The proposal reflects a broader recognition that agentic AI systems create security challenges that go beyond traditional model testing, and that solving them requires coordination across the industry.