The unprecedented breach of a technology vendor by an autonomous artificial intelligence agent serves to highlight a government-wide effort to streamline federal cloud security and prioritize faster patching of critical software vulnerabilities.
OpenAI admitted this week that its advanced AI training models were infiltrated from its own test environment and into the network of startup vendor Hugging Face. The case is the latest AI cybersecurity development to spur policy debate across Washington.
Pete Waterman, director of the General Services Administration’s Federal Risk and Authorization Management Program (FedRAMP), called the Hugface incident a landmark moment for cybersecurity at Carahsoft’s FedRAMP Summit on Thursday.
Waterman said the incident highlights his program’s transition to the FedRAMP 20x model.
“So when we talk about FedRAMP 20x and talk about changes to vulnerability management, vulnerability detection and response, etc., you have to understand that everything is going to be different,” Waterman said. “This is not a FedRAMP issue. This is not a compliance issue. If you think of FedRAMP as compliance, you’re done. You’re done. Get out of here. Your business will only survive if you integrate your security, engineering, and product teams to make changes at the pace of AI and fend off these attacks.”
The FedRAMP 20x initiative aims to reduce authentication time from years to weeks using automation, machine-readable data, and key security metrics. Waterman’s team began piloting this approach last year.
The program plans to stop accepting FedRAMP “Rev Five” certification packages and fully transition to the 20X model by next June.
Waterman said the government needs industry to focus on areas such as vulnerability detection and response, automation, integrating security into engineering and other security practices that are needed after an AI incident like a “face hug.”
“You shouldn’t do that for compliance,” Waterman said. “The compliance part of FedRAMP is a way to assure us that you’re doing that. But if your company isn’t willing to do it themselves, it takes three weeks to invest the right amount of money, put you in a part of the organization where compliance isn’t a department that everyone hates, and have a meeting with the engineering team there. It’s not going to be successful.”
“A new era of vulnerability management”
Federal agencies are also moving to implement faster, higher-priority software patching cycles based on the June Executive Order on AI Security and a corresponding binding operational directive from the Cybersecurity and Infrastructure Security Agency.
Under CISA’s directive, government agencies are now required to patch the highest-risk vulnerabilities on their networks within three days. However, agencies can significantly delay patching schedules for low-risk vulnerabilities.
Nick Polk, director of cybersecurity in the Office of the Chief Information Officer, said the Office of Management and Budget is working with CISA to ensure the directive is implemented across all departments.
“That can mean something as simple as making sure departments are accurately reporting to agency-level chief information officers and CISA,” Polk said at a July 16 event hosted by the Chamber of Commerce in Washington. “I can tell you that I have little patience for people who say, ‘Oh, I’m special. I can take care of myself. You don’t have to look at me at all.'” Of course, this doesn’t really work if component boundaries within an institution may mean a lot to the people in that institution, but they don’t mean much to sophisticated, persistent attackers who are willing to move seamlessly between those boundaries. ”
Will Rooks, senior director of intelligence in the Office of the National Cyber Director, said AI is accelerating the cycle of vulnerability discovery and exploitation, a cycle that has already accelerated in recent years.
“We’re just moving faster and faster through each stage of the cyber operations lifecycle that an attacker has to go through to get to a victim’s network and achieve an outcome,” he said.
In response to rising cyber AI risks, the Treasury Department also launched the “Gold Eagle” initiative, which aims to identify vulnerabilities in advanced AI models before they are exploited. A Treasury Department-led clearinghouse is coordinating the discovery of these vulnerabilities in collaboration with leading AI companies and other institutions.
According to the White House, the effort aims to “provide prioritized, actionable threat and remediation information to defenders across the federal government and private sector.”
Polk said CISA’s Continuous Diagnosis and Mitigation (CDM) program is critical to tracking how agencies prioritize and patch vulnerabilities across the government.
But he said agencies also need to make sure their systems are “properly mapped” to the CDM to fully understand the attack surface and prioritize vulnerabilities.
He said that while CISA can grade software vulnerabilities by risk from a general perspective, each agency needs to understand whether the vulnerability exists in its environment, where it exists, and how it may impact its mission.
“When you have this level of network awareness, attack service awareness becomes very important. This allows you to properly calculate risk at the speed of the machine,” Polk said. “You can’t just call someone and say, ‘I found 10.65.60. Is it yours? Can you tell me more?'” In this new era of vulnerability management, you can’t do that. ”
Copyright © 2026 Federal News Network. Unauthorized reproduction is prohibited. This website is not directed to users within the European Economic Area.
