AI Leaders Propose SAFE Guidelines for Cybersecurity Transparency
The Open Secure AI Alliance, comprising over 120 organizations, is developing new SAFE guidelines to enhance agentic AI cybersecurity, aiming to transform security incidents into shared protection for the entire ecosystem.
Intelligence analysis by Gemini 2.5 Flash
As the Black Hat conference begins, the Linux Foundation has released a Request for Comments on the Shared AI Findings Exchange (SAFE) guidelines. These guidelines, drafted by an Open Secure AI Alliance working group including NVIDIA, Cisco, and Amazon, propose a framework for confidentially collecting and analyzing AI security incidents to develop shared operating recommendations and…
Imagine smart computer programs, called AI agents, that can do many tasks, like helping you write or organize things. Just like you need to make sure your toys don't break or get lost, we need to make sure these AI agents are safe and don't accidentally do something wrong or share your secrets. A group of big tech companies is working together on new rules, called SAFE guidelines, which are like a shared playbook. If one agent has a problem, everyone learns from it so they can make all agents safer and more trustworthy, like sharing tips to keep all your toys in good working order.
Analysis
Forging a Collective Shield for Agentic AI
The Open Secure AI Alliance, a growing consortium of over 120 organizations, is taking a proactive stance on the cybersecurity challenges posed by agentic AI. Their latest initiative, the Shared AI Findings Exchange (SAFE) guidelines, represents a significant step towards fostering a collaborative defense strategy. Proposed by a working group that includes industry giants like NVIDIA, Cisco, and Amazon, SAFE aims to create a structured approach for handling AI-related security incidents. The core idea is to move beyond isolated responses to breaches and near-misses, instead transforming these events into actionable intelligence that can benefit the entire AI ecosystem. By confidentially collecting and analyzing incident data, informing affected parties, identifying recurring vulnerabilities, and publishing evidence-based recommendations, SAFE seeks to establish a virtuous cycle of learning and improvement, ultimately reducing systemic risk across AI deployments.
NVIDIA's Multi-Layered Security Contributions
NVIDIA is a key contributor to the Open Secure AI Alliance, offering a comprehensive suite of tools and models designed to secure the full AI stack. Their contributions span from research harnesses to runtime controls and model-level security. The NVIDIA Labs Object-Oriented Agent (NOOA) research harness, for instance, facilitates easier testing, tracing, auditing, and governance of agent behavior. Complementing this, NVIDIA OpenShell restricts agent actions, enforcing security and privacy controls at the agent level. Furthermore, NVIDIA's open model families, such as Nemotron for agentic AI and Isaac GR00T for robotics, are shipped with open weights and training techniques, promoting transparency. The company also provides open-source verified agent skills, which are cataloged, scanned for risks like prompt injection, cryptographically signed, and documented, ensuring defenders understand their capabilities and origins. Tools like NeMo Guardrails, NeMo Anonymizer, NeMo Safe Synthesizer, and the Garak LLM vulnerability scanner further bolster security by enforcing policies, protecting data, generating privacy-safe synthetic data, and identifying model vulnerabilities before deployment.
Broadening the Defensive Front with Alliance Partners
The collaborative spirit of the Open Secure AI Alliance extends beyond NVIDIA, with numerous members contributing to a robust, multi-layered defensive stack. These contributions address critical areas such as identity and permissions, harnesses, runtime guardrails, security AI models, observability, and data security. Okta, for example, is developing reference implementations for agent identity and access using the Cross App Access (XAA) protocol, enabling secure connections for agents in sandbox environments. Palo Alto Networks has contributed open-source tools like Agent Guard and Agent Watch from its Idira platform, helping developers apply identity security best practices. Red Hat's new open-source project, asago, maps organizational governance requirements to agent runtime controls, providing a single audit trail. Amazon, a recent addition to the alliance, contributes Strands Agents, an open-source toolkit for building transparent AI agents, and Cedar, an authorization language for enforcing verifiable boundaries on agent actions. This collective effort underscores the belief that shared threat intelligence and open, inspectable tools are essential for building a resilient and secure AI future.
Key points
- The Open Secure AI Alliance, with over 120 organizations, is developing SAFE guidelines for agentic AI cybersecurity.
- SAFE proposes a framework for confidentially collecting, analyzing, and sharing AI security incident findings to reduce systemic risk.
- NVIDIA contributes a full stack of open cybersecurity software, including NOOA, OpenShell, NeMo Guardrails, and the Garak LLM vulnerability scanner.
- Other alliance members like Cisco, CrowdStrike, Hugging Face, Red Hat, Okta, Palo Alto Networks, and Amazon are also contributing tools and frameworks.
- The initiative emphasizes collective defense and open, inspectable tools to strengthen AI security across identity, harnesses, runtimes, and data privacy.
The collaborative development of SAFE guidelines and shared security tools by a broad alliance of AI leaders could significantly accelerate the establishment of robust, standardized cybersecurity practices for agentic AI. This collective defense approach promises to enhance trust, foster safer AI deployment, and reduce systemic risks across the rapidly evolving AI landscape.
Despite the collaborative efforts, the rapid pace of AI innovation and the continuous emergence of new attack surfaces could make it challenging for security guidelines and tools to keep up. Slow adoption rates or insufficient enforcement across the diverse ecosystem might also limit the overall effectiveness of these initiatives, leaving parts of the AI infrastructure vulnerable.
%20China-Free%20Robot.jpg)
