Tenable and OpenAI Launch AI Inspector for Secure Open-Source AI Agents
Tenable and OpenAI have introduced the CyberAgents Exchange AI Inspector, a new tool designed to rigorously vet open-source AI agents before they are listed on the CyberAgents Exchange registry.
In a significant move to bolster the security of the burgeoning open-source AI ecosystem, Tenable and OpenAI have unveiled the CyberAgents Exchange AI Inspector. This new tool is integrated into Tenable's CyberAgents Exchange registry, a platform dedicated to hosting AI agents, skills, MCP servers, and multi-agent playbooks. The Inspector aims to provide a rigorous and transparent security review process for all community-submitted AI agents, ensuring a higher standard of safety and trustworthiness for users.
The CyberAgents Exchange AI Inspector combines Tenable's extensive expertise in exposure management with OpenAI's advanced GPT Cyber models. This powerful synergy allows for a comprehensive analysis of an AI agent's attack surface, which extends beyond traditional software vulnerabilities. Key areas of focus include the security of Large Language Model (LLM) instructions, the permissions granted to tool-chaining capabilities, and the potential risks associated with prompt injection attacks. Human oversight is also a critical component of the review process, ensuring that complex or novel threats are adequately addressed.
AI agents, by their nature, present a unique set of security challenges. Unlike conventional software where the attack surface is primarily the code itself, AI agents interact with LLMs, external tools, and potentially sensitive data. The Inspector is designed to account for this expanded threat landscape. It examines not only the code but also the instructions given to the AI, the tools it can access, and the data it can process or transmit. The risk of prompt injection, where malicious inputs can manipulate the AI's behavior, is a particular concern that the Inspector actively seeks to mitigate.
Submissions to the CyberAgents Exchange are anchored to specific commits in their respective code repositories. This ensures that findings from the Inspector are traceable to a known, static state of the agent. Any significant changes to a submission necessitate a new review, maintaining the integrity of the vetting process. Furthermore, the Inspector assesses the trustworthiness of repository owners and associated organizations, using various signals to generate a confidence rating that informs the depth of the review.
The vetting process itself is multi-layered. It begins with Tenable One AI Exposure scanning, which automatically identifies potential issues like Personally Identifiable Information (PII) exposure, unauthorized sensitive data access, and prompt injection or jailbreak attempts. Submissions that pass this initial automated screening then undergo frontier model assessment using OpenAI's GPT Cyber models. Finally, human reviewers provide an additional layer of scrutiny. The CyberAgents Exchange also installs and exercises submissions in an isolated environment to compare observed behavior against documented claims, flagging any discrepancies.
Findings are categorized based on severity. Hard security failures must be resolved before an agent can be listed. Other, less critical findings may be accepted with a documented rationale and a commitment to future remediation. Tenable emphasizes a collaborative approach, working directly with submitters to improve their agents rather than operating as a purely restrictive gatekeeper. The goal is to foster the adoption of secure AI technologies.
The CyberAgents Exchange, launched in August, has already amassed over 100 AI listings. The introduction of the AI Inspector marks a significant enhancement to its commitment to security. By providing a robust review mechanism, Tenable and OpenAI aim to empower security teams to adopt agentic AI more quickly and with greater confidence, paving the way for more secure and reliable AI-powered applications.