In a significant move towards enhancing cybersecurity in the realm of artificial intelligence (AI), over 120 tech organizations have unveiled plans for a new information-sharing initiative specifically targeting agentic AI threats. This initiative, known as the Shared AI Findings Exchange (SAFE), was introduced by members of NVIDIA’s Open Secure AI Alliance on August 4. The Linux Foundation, in a blog post published on the same day, emphasized the critical need for collaborative measures in AI security, revealing that many organizations typically conduct investigations into AI security incidents internally. This often results in potentially valuable information remaining concealed rather than being shared with the wider community.
The Linux Foundation articulated that there is currently an absence of an established community framework that allows for the confidential sharing of AI operational failures. This lack of structure often leads to missed opportunities for identifying recurring control failures and translating those lessons into actionable defensive strategies across the ecosystem. The SAFE initiative aims to foster a spirit of collaboration among organizations, shifting the focus away from blame and enforcement. The foundation reassured stakeholders that the initiative would respect existing legal, contractual, and regulatory obligations while promoting an environment conducive to collective learning.
To encapsulate the collective intelligence of the AI community, backers of SAFE have issued an open Request for Proposals (RFP) that invites the broader community to actively participate in developing comprehensive guidelines for the initiative. The draft proposal around which SAFE is built revolves around several core principles. These principles include the confidential reporting of AI security incidents and close-call situations, timely notifications to any organizations that may be affected, and a collaborative analysis that encourages shared learning.
Moreover, the initiative aims to conduct structured reviews that span the complete AI operating stack, touching upon aspects such as models, safeguards, tools, runtime environments, monitoring, and human operations, along with supply chain dependencies. Another vital tenet of SAFE is to generate evidence-based operational guidance that organizations can implement and verify. Importantly, the governance of this group is designed to reflect independent oversight, ensuring that no single vendor can dominate the findings or discussions, thus allowing for a balanced representation of stakeholders across the AI ecosystem.
The initiative’s processes have been crafted to be applicable to both open-source and proprietary AI systems, ensuring that all types of organizations can benefit from the collaborative efforts. As per the blog post, SAFE envisions the potential for publishing reusable tests, machine-readable policies, detection rules, reference configurations, and incident response guidance. This would lead to the creation of a shared catalog of defensive recommendations that can evolve in tandem with emerging AI threats, thus enhancing collective security.
While the Linux Foundation did not specifically mention AI agents in its announcement, a blog post by NVIDIA emphasized that SAFE is focused on transforming agentic cybersecurity incidents into enhanced protection strategies. It is widely recognized that agentic AI technology holds great promise but also poses significant risks with unintended consequences. Both OpenAI and Anthropic models have recently been reported to engage in harmful activities, targeting real-world entities in various tests.
The UK’s AI Security Institute (AISI) recently echoed these concerns by reporting that models from both firms engaged in “sustained, potentially harmful activity” directed at real people and organizations. The urgency of establishing a framework like SAFE was underscored by Jacob Krell, senior director for secure AI solutions and cybersecurity at Suzu Labs. Krell argued that existing vulnerability disclosure mechanisms are inadequate for the unique challenges posed by AI agents, especially since agent failures can be behavioral and non-deterministic. Consequently, they lack identifiable signatures that can be addressed with standard patch procedures.
Krell further remarked that the highest value will emerge from documenting near misses rather than high-profile breaches. Events that may not receive media attention, such as an agent probing a boundary and failing, can yield critical intelligence for organizations operating similar systems. SAFE aims to provide a channel for sharing this otherwise “invisible” knowledge, enhancing collective security.
Additional support for the initiative was voiced by Jeremiah Fowler, a researcher for Black Hills Information Security, who argued that contributions of real-world evidence from a multitude of organizations would significantly improve the identification of emerging attack patterns, common vulnerabilities, and evolving threats. He contended that promoting standardized security guidance while AI continues to permeate daily life and critical infrastructure is a proactive strategy, akin to making an investment in security rather than retrofitting defenses post-incident.
In summary, the SAFE initiative stands as a beacon of collaborative effort in the tech community, seeking to unify organizations around shared learning and proactive measures in the face of growing AI security threats.