Vulnerability Management

The Trump administration announced the launch of a program that will coordinate the security community’s use of frontier AI models to rapidly identify and fix vulnerabilities.

The vulnerability management clearinghouse, which the administration is calling Gold Eagle, is a response to the skyrocketing number of vulnerabilities that AI models are finding and the strain that that surge is placing on the security community. Through the clearinghouse, the government will coordinate efforts by private companies and independent researchers to scan critical software packages for flaws, fix any flaws that surface and deploy those fixes to end users.

With so many security experts looking for software vulnerabilities, the government is trying to marshal their expertise — and the capabilities of the frontier AI models that many of them are using — in a harmonized way. The goal is to deploy the nation’s vulnerability hunters and their AI tools across the widest possible range of software, rather than see some of them duplicate their efforts by focusing on the same software, squandering precious time and resources in the arms race against cybercriminals and nation-state adversaries.

Gold Eagle “has already begun to intake and prioritize identified cybersecurity vulnerabilities from across industries and sectors, coordinate scanning verifications, and ultimately ensure the security of our nation’s software and networks,” the White House said in a statement.

The core of the Gold Eagle program is the Vulnerability Information and Coordination Environment (VINCE), which the government is operating in partnership with Carnegie Mellon University’s Software Engineering Institute. The VINCE platform will allow anyone to report vulnerabilities to the Gold Eagle program for triage and mitigation.

VINCE will enable “vulnerability and patching coordination at a speed and scale never seen before,” National Cyber Director Sean Cairncross told reporters during a briefing.

Gold Eagle will focus heavily on open-source software, whose code underpins a wide range of critical infrastructure but is often poorly scrutinized. Open-source developers, many of whom volunteer their time to maintain their projects, say they have been overwhelmed by a tidal wave of AI-generated bug reports, some of which are startlingly accurate.

Open-source developers “are important partners” in the Gold Eagle program, Cairncross said, describing their code as essential to American life.

Redundancy, liability concerns

The White House described Gold Eagle — the creation of which President Donald Trump mandated in his June executive order on AI security — as “a coordinated system to receive and patch cyber vulnerabilities at a speed and scale never seen before,” one that “represents a new operational model for cyber defense.”

Even so, Gold Eagle arrives at a time when the private sector has already created several similar programs. The Linux Foundation, with support from Anthropic, Microsoft and other leading tech companies, has created a program called Akrites to boost the open-source community’s identification and handling of vulnerabilities. And the open-source security vendor Chainguard has partnered with Cisco, Cloudflare, JPMorgan Chase and other major businesses to launch Athena, another vulnerability coordination system focused on open-source software.

Those two industry-led programs involve frontier AI firms and participants in Anthropic’s Project Glasswing and OpenAI’s Daybreak coalitions. The Trump administration has not identified the companies participating in the Gold Eagle program, including which AI firms are contributing their resources to the effort. (Anthropic has said that it would participate in the government-led clearinghouse.)

One potential obstacle already looms large over Gold Eagle: Its system of exchanging vulnerability information relies on the liability protections of the Cybersecurity Information Sharing Act, which Congress in February temporarily reauthorized through the end of September. The Trump administration has called on lawmakers to reauthorize the program for 10 years, saying its protections are essential to a robust cybersecurity collaboration ecosystem.

For decades, Cisco has positioned itself as a leader in network infrastructure. The company is one of the top providers of secure networking equipment that businesses and governments depend on. In recent months, that dominance is proving to be the company’s Achilles’ heel. 

Since early 2026, security researchers have been tracking an ongoing series of attacks targeting vulnerabilities in Cisco SD-WAN. Those attacks appear to have impacted key government sites and critical infrastructure providers and have raised the anxiety level of authorities across the globe.

“Edge infrastructure remains one of the highest-value targets in enterprise security,” said Douglas McKee, director of vulnerability intelligence at Rapid7. “If you compromise SD-WAN or firewall management, you’re landing on policy, visibility, routing, segmentation, and, in many cases, administrative trust over a large swath of the environment.”

Cisco environments have become an increasingly important target by state-nexus and other top threat actors in recent years, due, in part, to the wide use of these tools in the most sensitive government and enterprise networks in the world.

During the 2024 campaign by China-nexus actor Salt Typhoon, the attacks leveraged access to Cisco devices to reach into major telecommunications providers.

Threats to critical sectors

The Cybersecurity and Infrastructure Security Agency in February issued an emergency directive warning that a threat actor was targeting an authentication bypass vulnerability, tracked as CVE-2026-20127, and a privilege-escalation flaw, tracked as CVE-2026-20775, in Cisco Catalyst SD-WAN systems. 

Cisco Talos researchers in May said an authentication bypass vulnerability, tracked as CVE-2026-20182, was being targeted by a sophisticated threat actor tracked as UAT-8616. 

Other threat actors chained together three vulnerabilities in Cisco Catalyst SD-WAN Manager, which allowed attackers to gain access to the device, Cisco Talos said. The chained vulnerabilities, CVE-2026-20133, CVE-2026-20122 and CVE-2026-20128, led to the release of web shells that allowed attackers to execute bash commands on a device.

Cisco products are frequently targeted due to their widespread use in government and enterprise networks, Jonathan Forest, a VP analyst at Gartner, said. 

A software-defined wide-area network is a virtual technology used to connect various business locations, including cloud environments, data centers and branch locations. In an SD-WAN environment, users get a combination of efficient routing, security and load balancing. 

Cisco Catalyst SD-WAN is often implemented in a customer’s own environment, which means customer security teams are largely responsible for patching their own software. 

“The result can be delays in patching and vulnerabilities being exposed for an extended time, leaving it open for attackers,” said Forest. 

Because Cisco products are so widely deployed, there’s a lot riding on their resilience and security.

“Due to that prevalence, their products often come under increased attack, since a single flaw can result in access to tons of downstream high-profile customers,” said TJ Sayers, senior director of threat intelligence at the Center for Internet Security. 

Critical services

Hospitals and other healthcare facilities are highly dependent on technology provided by SD-WAN. They often rely upon distributed networks where a central hospital is connected to an off-site health clinic or data center, or use a cloud platform. 

“When critical vulnerabilities emerge repeatedly in wide-area network infrastructure, it puts immense pressure on health sector IT security teams and the stakes are high,” Errol Weiss, chief security officer of the Health Information Sharing and Analysis Center, told Cybersecurity Dive.

“If a cyberattack leverages an infrastructure flaw to cause a prolonged IT outage that disrupts emergency care, lab results, or medication management, care slows down, negatively impacting patient outcomes,” Weiss said. 

Unauthorized peering

Mandiant, the incident response arm of Google Cloud, disclosed a March attack where a hacker exploited a zero-day vulnerability in Cisco Catalyst SD-WAN at a communications service provider to gain root-level access through a compromised administrative account. The vulnerability was later identified as CVE-2026-20245 and a patch was released. 

Mandiant initially observed unauthorized peering connections to the service provider’s SD-WAN devices starting in late 2025, noting the possibility the hacker exploited CVE-2026-20127 or CVE-2026-20182, which had not been disclosed or patched at the time. It was later determined that the targeted device was not affected by either of the prior vulnerabilities. 

Researchers said the March incident also involved unauthorized peering, which is the process of establishing a trusted digital connection between distinct network components, such as edge routers, regional hubs or central controllers. After gaining access, the hacker authenticated to the SD-WAN Manager and changed the password on the default administrator account. 

Mandiant researchers said the hacker took numerous steps to cover up their forensic footprint, including deletion of all files created during the attack and restoring system configuration that were modified. 

It is not immediately clear whether the hacker in the late 2025 incident and the March incident are the same. 

“Mandiant suspected it might be a zero-day during its initial investigation, but deeper forensic analysis was required to confirm the exploit of a new flaw,” Pete Boonyakarn, senior cybersecurity consultant at Mandiant - Google Cloud, told Cybersecurity Dive.  

Highly targeted

The attacks against SD-WAN are not an entirely unique experience for Cisco, as the company’s networking and security products have been among the most targeted in recent years. 

The Cybersecurity and Infrastructure Security Agency issued an emergency directive in September 2025 ordering federal civilian executive branch agencies to take immediate action to mitigate several vulnerabilities in Cisco devices. 

The directive was linked to a sophisticated actor exploiting zero-day flaws in Cisco Adaptive Security Appliances and then further manipulating read-only memory to enable persistence that survived reboots. The same flaws were also discovered in Cisco Firepower devices. 

At least 10 organizations globally had been compromised at the time of the disclosure, and that number was expected to increase over time. 

The attacks were linked to a remote code execution flaw, tracked as CVE-2025-20333, and a privilege-escalation flaw tracked as CVE-2025-20362. The threat activity linked back to the Arcane Door campaign that began in early 2024. 

The September incident drew the attention of Congress. In October 2025, Louisiana Sen. Bill Cassidy sent a letter to Cisco asking a list of specific questions about the potential impact on federal agencies and the larger public. 

Cassidy, chair of the Senate Health, Education, Labor and Pensions Committee, said he was leading an investigation into the potential impact of cyber risk calling it a “grave national threat.” 

“As the largest provider of network infrastructure in the world, Cisco holds a unique position in delivering not only to the federal government, but virtually all businesses,” Cassidy wrote in the letter. 

Security governance

Despite its recent history of issues, Cisco is widely considered to be deliberate and thoughtful about the security of its products and the trust of its customers. 

The company operates a Security and Trust Organization under the leadership of Chief Security and Trust Officer Anthony Grieco, according to the 2025 Cisco annual report. The company board of directors monitors cyber risk through an audit committee, and the committee meets with Grieco multiple times per year. 

The board of directors and audit committee get more frequent updates in the event of cybersecurity incidents or threats. At the time the annual report was filed, Cisco did not believe its financial condition or operating results were materially impacted by cyber risk or any previously identified cyber issue.

Proactive measures

Cisco executives publicly acknowledged the rapidly changing threat landscape in recent years, and the company has programs designed to maintain the security of its product line.  

“We have a red team that goes off and attacks our products proactively, trying to find vulnerabilities,” Grieco said during a panel discussion at the Cisco Live conference in Las Vegas.

However, Cisco executives understood they could not afford to rely on traditional measures to stop modern day threats. Cisco has recently embraced frontier AI models to push the boundaries of vulnerability scanning to a scope and scale the best human testers cannot achieve alone. 

Cisco was a founding member of Project Glasswing and participated in private testing of Anthropic’s advanced Claude Mythos Preview. 

Cisco used frontier AI to test 1.8 billion lines of code over a period of eight weeks, Grieco said in a June blog post. He noted the work would have taken the company’s security team about eight years to complete such a task. 

Cisco officials understand that AI alone will not magically resolve all of their product vulnerabilities. Grieco said the company combined frontier LLMs with its own human-guided harness and was able to achieve a false positive rate of less than 3%. 

“True AI-driven security is measured by actionable precision at scale, not by the count of vulnerabilities alone,” Grieco noted in the blog post. 

The company has unveiled key changes in how it announces and prioritizes vulnerability disclosure. 

Starting in July, Cisco moved to a twice-monthly disclosure model, beginning on the first and third Wednesday of each month, according to a blog post from Russ Smoak, Cisco’s vice president of information security. 

The company has also shifted its disclosure model to emphasize critical vulnerabilities that are actively exploited or flaws that pose the highest risk of becoming exploited, Smoak said.

Sponsored content

By Tanium

Ask most CISOs what worries them today, and “we don’t know where our risk is” is rarely the answer. Visibility has gotten good. Vulnerability scanners, asset inventories, and exposure management platforms can produce a prioritized risk list by lunchtime. The harder question, and the one that actually determines whether an organization stays ahead of a breach, is what happens after that list gets produced.

That gap between identification and remediation is now the defining challenge in exposure management, and it’s backed by data, not just anecdote. A recent IDC PeerScape report on exposure management best practices, built on interviews with security leaders across financial services, manufacturing, healthcare, education, and technology, found that fewer than half of organizations use real prioritization algorithms: most still lean on CVSS scores that don’t reflect their actual environment or business context. The same research points to asset ownership delays as a leading driver of slow patch cycles, and recommends a mindset shift: the goal isn’t eliminating every exposure, it’s reducing the attack surface that matters most, faster.

The Math Is Shifting Back in Defenders’ Favor

The traditional lifecycle (discover, assess, prioritize, remediate, verify) was built for a world where threats moved at human speed. Attackers are now using frontier AI models to find and weaponize vulnerabilities in a fraction of the time it used to take. Meanwhile, remediation inside most organizations still runs through tickets, spreadsheets, and handoffs between vulnerability management, IT operations, and security operations teams that don’t share a single view of the problem. Even the strongest detection only pays off once remediation moves at the same speed.

This isn’t primarily a data problem. Most teams have plenty of data, often more than they can use, spread across tools that don’t yet talk to each other. The constraint is operational: turning a validated finding into a completed action at the speed the threat landscape now demands.

Where AI Actually Changes the Equation

Most AI applied to exposure management so far has sharpened the “find” side: smarter risk scoring, predictive exploitability, natural-language queries. Useful, but it optimizes a step that was already comparatively fast. The bigger opportunity is applying AI to the parts of the lifecycle that have resisted automation: triage, orchestration, and remediation itself.

This is the shape agentic AI is starting to take across the industry: systems that observe an environment continuously, plan a fix grounded in real-time data, and execute multi-step remediation across a fleet, while still pausing for human approval before anything consequential runs. That last part isn’t a limitation; it’s the design principle that makes automation trustworthy. Closing the loop responsibly means keeping people in control of judgment calls, while removing the manual friction around everything that leads up to them.

The New Standard for Security Leaders

Exposure management is moving past measuring itself by how well it reports on risk. Reporting is the starting point, not the finish line. The programs that hold up under pressure are the ones that compress the entire distance between “we found it” and “it’s fixed.” Increasingly, that compression happens fastest when AI is doing the heavy lifting on triage, orchestration, and remediation, not just detection.

That’s the real inflection point for security leaders right now. The organizations that treat exposure management as one continuous, AI-driven motion, rather than a chain of tools stitched together by tickets, will close exposures faster. They’ll close the door before attackers know it was open.

In the weeks before technology vendors disclose new software vulnerabilities, hackers sometimes stumble upon the flaws and begin exploiting them prior to customers even knowing there’s a problem.

In a report, the internet intelligence firm GreyNoise revealed that roughly half of the scanning and exploitation activity surges it tracked between mid-December 2025 and late March 2026 were followed, within the next three weeks, by vulnerability disclosures from the targeted vendors.

Nearly two-thirds of the activity surges led to vulnerability disclosures within six weeks, according to the report.

“Scanning and exploit activity targeting specific vendors consistently rose before those same vendors disclosed new CVEs,” GreyNoise said.

There was a surge of exploitation of a severe Cisco vulnerability — one that prompted a rare emergency directive from the Cybersecurity and Infrastructure Security Agency (CISA) — as early as 39 days before Cisco disclosed the flaw, according to the report. A similarly critical VMware vulnerability saw exploitation 36 days before disclosure, and a major MikroTik vulnerability saw exploitation 24 days before disclosure. GreyNoise also discovered similarly early exploitation of other high-severity flaws from Juniper, SonicWall and Ivanti.

In the case of the Cisco flaw, GreyNoise saw five exploitation surges over the final 18-day period preceding disclosure, with the number of IP addresses plummeting between surges, even as the number of sessions increased — a phenomenon GreyNoise said was “consistent with a shift from broad reconnaissance to dedicated operators hammering specific targets.”

One potential bright spot emerged from the data: “The highest-severity threats tend to generate substantial probing activity and meaningful lead times,” according to the report.

GreyNoise gathered its data by analyzing scans, brute-force login attempts, remote-code-execution probes and other attacks against products from 18 edge device and network infrastructure vendors over a 103-day period. The security firm evaluated both the volume (in terms of unique sessions) and the breadth (in terms of unique IP addresses) of the activity. Each multiday period of above-average activity targeting a specific product constituted a “spike event.”

An early warning

The median amount of time between a surge of exploitation and a vulnerability disclosure was 11 days, which GreyNoise noted could be a significant “head start” for companies that learned about the surge. (GreyNoise sells threat intelligence that includes information about such activity.) “Eleven days is enough time to brief leadership, stage a patch, and harden exposed systems before the rest of the world learns the vulnerability exists,” the company said.

Not all activity is equally predictive

GreyNoise’s report breaks down the type of activity that it observed in each surge and how often the targeted vendor later disclosed a vulnerability. The company saw 42 instances of scanning, with 57% of them leading to vulnerability disclosures; 18 brute-force attempts, with 56% leading to disclosures; and 12 attempts to execute remote code, with 42% leading to disclosures.

The techniques were associated with different lead times — scanning typically occurred further back from vulnerability disclosure than brute-force and remote-code-execution attempts, which GreyNoise noted was “consistent with later-stage activity, where attackers have already identified their targets and are trying to get in.”

Scanning was also more likely to be widely dispersed across IP addresses, with each one responsible for only a few sessions, whereas later-stage activity was more concentrated, with a small number of IP addresses each registering a large number of sessions.

Less than 1% of software vulnerabilities were exploited in the wild over the past year, but those flaws are being weaponized faster and on a larger scale than ever before, according to a report by VulnCheck. 

Researchers tracked more than 14,400 exploits linked to about 10,500 unique CVEs in 2025, representing a 16.5% increase from the prior year. A large percentage of that increase was linked proof-of-concept code that was generated by AI. 

Researchers cautioned that much of that AI-generated code was non-functional. 

The results highlight the difficulties security teams have in prioritizing which are the most serious threats that need to be investigated. Threat groups are increasingly able to weaponize flaws before network defenders can apply security patches and take other mitigation measures.

“Defenders have long taken the availability of public exploit code as a significant risk signal,” Caitlin Condon, vice president, security research at VulnCheck, told Cybersecurity Dive. 

Condon said the research shows that an overwhelming amount of AI-generated information is creating problems for defenders trying to judge what are legitimate threats and what can be ignored.

More than 50% of CVEs linked to ransomware were first identified as a result of a zero-day vulnerability.  

React2Shell, tracked as CVE-2025-55182, was the top vulnerability of 2025, with 236 known exploits. 

A Microsoft Sharepoint vulnerability, tracked as CVE-2025-53770, had 36 known exploits.

AI models are helping discover software vulnerabilities even faster than previously expected. 

Lee Klarich, the chief product and technology officer at Palo Alto Networks, in a blog post shared PAN's insights into how frontier AI models are impacting cybersecurity. 

Palo Alto Networks is among a handful testing Anthropic’s Claude Mythos Model since April 7 as a launch partner under Project Glasswing. Beyond Claude Mythos, the company has also been testing Claude Opus 4.7 and OpenAI’s GPT 5.5-Cyber, according to Klarich. 

Palo Alto Networks released its first set of Patch Wednesday security advisories and disclosed 26 common vulnerabilities and exposures, compared to its usual volume of about five, Klarich said, noting that none of the newly disclosed vulnerabilities are currently being exploited in the wild. 

Klarich cautioned that simply running one of the models would not automatically resolve the vulnerability problem. Organizations need to build “AI scanning harnesses, leverage context, guardrails and threat intelligence” to successfully find and remediate these flaws at scale. Security teams also should develop a “multimodal approach,” he said, in order to identify a superset of vulnerabilities. 

The longer-term play would be to incorporate these models further into the software development life cycle. “This is the light at the end of the tunnel,” Klarich said in the blog. “A future where software is secure by design.”

He warned that organizations have a narrow “three- to five-month window” for organizations to get a leg up on adversaries, as AI-driven exploitation will likely become the new normal. 

Researchers at Google Threat Intelligence Group reported that AI had been used to develop a working zero-day exploit. Microsoft said it found 16 new vulnerabilities, including four critical remote execution flaws, across the company’s networking and authentication stack.  

A threat actor was able to leverage AI to develop a working zero-day exploit, in what is believed to be the first such successful use of the technology, according to a report by Google Threat Intelligence Group (GTIG). 

The effort was an attempt to launch a mass exploitation event, the report read, but ultimately was unsuccessful, as Google discovered it before the vulnerability was weaponized. 

GTIG notified the developer of the exploit and a patch was issued to address the potential threat. Researchers do not believe that Mythos was used in the development process. 

“AI can review the underlying logic, context, and flow of code at scale to discover vulnerabilities,” John Hultquist, chief analyst at GTIG told Cybersecurity Dive, via email. “It can also be used to build working exploits which are a significant hurdle.”

The incident highlights a growing trend by state-linked and financially motivated threat groups using AI to scale and accelerate hacking campaigns and exploit flaws in widely used applications. GTIG researchers caution they have seen several other attempts to use AI to develop exploits and expect there will be other operations employing a variety of models. 

Threat groups affiliated with North Korea and China have shown significant interest in exploiting AI to discover vulnerabilities, according to the GTIG report. 

For example, a North Korea-linked hacker tracked as APT45 has used AI to analyze a wide range of vulnerabilities using thousands of repetitive prompts and validate proof of concept exploits, the report stated.  

In a recent case involving criminal actors, a group of hackers joined forces to plan a mass exploitation operation. A zero-day vulnerability was implemented in a Python script, which enabled the hacker to bypass two-factor authentication on a widely used open-source system administration tool, GTIG said. 

Researchers worked with the vendor to disclose the vulnerability and disrupt the operation. 

The report follows an attempt by an unknown hacker to breach a Mexican water utility using widely used AI tools, including Claude.