Nvidia Backed New Exchange to Track AI Agent Failures

The Secure Agent Findings Exchange aims to standardize failure reporting across 140 firms to address security gaps.

Updated on Sept. 20, 2026 in Artificial Intelligence

Isometric editorial illustration of metallic cooling fins and geometric paths representing digital execution traces for software security. — Nvidia and a 140-firm coalition have launched the Secure Agent Findings Exchange to standardize the tracking and diagnosis of autonomous AI coding failures. AI Illustration. Upload story photo >

Live Poll

Should companies be required to share internal data about AI system failures with other organizations?

Nvidia has announced support for the Secure Agent Findings Exchange (SAFE), a new industry initiative backed by approximately 140 companies that creates shared infrastructure for reporting agent failures. The project seeks to address the issue of coding agents failing on more than 60% of tasks.

Why it matters

Sharing failure data prevents companies from repeatedly discovering the same agent vulnerabilities, which are difficult to diagnose because agents can change course during execution. This initiative provides a unified framework for identifying the root causes behind persistent agent errors.

Monitoring persistent agents currently adds 20% to inference compute requirements. Nvidia is promoting the use of its OpenShell runtime for sandbox isolation and policy enforcement to manage these agent stacks.

The players

Nvidia

A semiconductor and software company that designs GPU hardware and software stacks for training and running large-scale AI models.

CrowdStrike

A cybersecurity firm that builds endpoint protection and threat intelligence software currently integrating AI models to automate security tasks.

The details

Nvidia classifies the modern agent stack into three layers: model, harness, and runtime. The exchange allows developers to reconstruct agent execution traces—a chronological record of software operations—to identify the exact point of failure. Organizations can then utilize this shared reporting infrastructure to distribute information about discovered vulnerabilities. Additionally, companies like CrowdStrike are fine-tuning Nemotron models to create paired security agents capable of monitoring these execution flows.

September 20, 2026: The initiative was formally announced.

The Tech Race

This initiative attempts to replicate the collaborative vulnerability reporting model established by the CVE system for the emerging domain of autonomous AI agents. It represents a pivot from individual security efforts toward a centralized framework to stabilize the agent-based software ecosystem.

For developers, this exchange will eventually simplify debugging by providing a database of known failure patterns for AI coding tools. These monitoring tools carry an overhead cost of 20% in inference compute, which may influence how organizations decide to deploy agents in production.

The takeaway

The move toward a shared registry suggests that AI reliability is moving from a proprietary concern to an industry-wide security priority. Watch for updates from the participating 140 companies on the specific data schemas they will use to contribute to the exchange.