Nvidia suggested artificial intelligence agents should be equipped with digital black boxes that record every action they take.
Justin Boitano, the vice president and general manager of enterprise computing at Nvidia, claimed the Shared AI Findings Exchange (SAFE) program is based on NASA’s Aviation Safety Reporting System, speaking to Axios at the Black Hat USA 2026 conference. A tech coalition of more than 120 companies, including Nvidia, Hugging Face and CrowdStrike, proposed SAFE to create a flight-recorder-style system for reporting AI agent incidents on Aug. 4.
“The way I think of it is the harness, which has visibility into everything the agent is doing, is the flight recorder,” Boitano told Axios. “If you can get cybersecurity experts access to the flight recorders when these accidents happen, they can make a better determination on the right set of controls for the industry.”
SAFE would serve as an independent system to report AI security issues, according to the proposal.
Among the representatives for the proposed security apparatus are AI model developers, independent security and safety researchers, and government bodies. The government would act as “non-controlling observers,” the proposal stated.
“The [SAFE] draft proposal is intended to start an open community discussion around how organizations can confidentially learn from AI security incidents and near misses, transforming those experiences into practical guidance that strengthens security across the broader AI ecosystem,” the Linux Foundation said in the Aug. 4 announcement.
The proposed “notification timelines” would have SAFE members alert affected organizations right away, file a confidential report with SAFE within four business days, release an initial report within 30 days when appropriate and share weekly updates on unresolved material risks, the proposal said.
Nvidia and the Linux Foundation each did not immediately respond to the Daily Caller News Foundation’s requests for comment.
“Cybersecurity is strongest when we work together,” Sagar Dahatonde, Engineer II at CrowdStrike, said in an X post expressing CrowdStrike’s support for the SAFE program.
SAFE offers no formal legal protections for companies that report AI security incidents voluntarily, Axios reported.
“There’s been very little pushback,” Deputy CISO and Vice President at Nvidia, Julien Soriano, told Axios. “We see people wanting to get on board. They want to share.”
OpenAI admitted that its advanced AI models, including GPT 5.6, escaped a controlled test and hacked AI startup Hugging Face, according to OpenAI’s July 21 announcement. Meanwhile, Anthropic’s AI models circumvented a capture-the-flag testing environment and acquired unauthorized access to three unspecified organizations, according to Anthropic’s July 30 announcement.
An AI agent is an autonomous system that can create a series of steps, or workflows, to complete tasks on its own, according to IBM. Compared to regular AI models that primarily answer prompted questions, AI agents can plan, use tools, and take action on their own to complete goals.
(DCNF)
