Share on social
Nov 18, 2026
Lorem ipsum
.jpg)
When Anthropic announced Project Glasswing in April 2026, it brought attention to the reality that frontier AI models were now capable of identifying serious vulnerabilities in widely used software at scale, including flaws that had remained undiscovered for years.
The underlying capability is significant. Models such as Claude Mythos or GPT-6 demonstrate how AI can radically accelerate vulnerability research, analyze large codebases and uncover weaknesses that might otherwise take experienced researchers considerable time to find.
But several months on from Project Glasswing’s inception, an analysis by VulnCheck researcher Patrick Garrity raises questions about how effectively that discovery capability has translated into security outcomes. According to Anthropic's figures, its models had identified 37,966 potential vulnerabilities by November 2, but only 7,349 of those findings (19.4%) appeared in its public disclosure ledger. Of the entries in that ledger, Garrity found 217 marked as fixed, while 1,654 were recorded as duplicates and another 419 had been withdrawn. The gap between discovery and remediation had also widened since his September analysis: more than 4,600 additional findings had been made public, but the number marked as fixed had increased by just 15.
What is the value of discovering vulnerabilities at unprecedented scale if the processes for validating, disclosing and fixing them cannot keep pace?
AI is making it easier to find vulnerabilities, but it is also exposing the limitations of the processes that turn research into better security. Discovery is accelerating while validation, disclosure and remediation remain constrained by human capacity and established workflows. The result is more findings, but not necessarily a corresponding improvement in security.
That’s not going to make the job for organizations already struggling to understand and reduce their exposure any easier.
AI is changing the economics of vulnerability discovery
The capabilities demonstrated by Mythos and other frontier models are part of a broader shift in vulnerability research. Models can analyze source code, identify suspicious patterns, generate and test hypotheses, and work through complex vulnerability chains at a speed that would be difficult to replicate through manual investigation alone.
This does not mean that AI has made vulnerability research an entirely automated process. The most effective use of these models still involves experienced researchers directing the investigation, providing context and determining whether a potential finding is genuine. But it does mean that researchers can cover more ground, investigate more deeply and pursue lines of inquiry that might previously have been too time-consuming.
The implications extend beyond defensive research teams. The same advances are available to attackers, who can use AI to accelerate reconnaissance, identify weaknesses and develop exploit techniques. As the cost and effort involved in vulnerability discovery and exploit development fall, the speed and scale of attacks increases.
Finding vulnerabilities is only half the battle
A vulnerability finding is the beginning of a process, not the end. It needs to be reproduced and validated, its impact understood, and its severity assessed in context. It then needs to be responsibly disclosed, investigated by the relevant vendor or maintainer, and ultimately addressed through a patch or mitigation that can be deployed by the organizations using the affected software.
Each of those stages takes time and expertise, putting pressure on the limited resources of software maintainers. As the volume of AI-generated findings increases, software maintainers face the prospect of more reports to investigate without a corresponding increase in the resources available to assess them. If validation capacity cannot keep pace, the most important vulnerabilities could become harder to identify and address.
The ability to generate findings and the ability to turn them into meaningful security improvements are two different things, and that applies both to the maintainers of software and the organizations downstream that rely on that software. A model might identify a serious vulnerability in a widely deployed product, but that does not tell an organization whether the affected software is part of their attack surface, whether the vulnerable version is exposed, whether the issue is exploitable in its environment, or whether attackers are already targeting it.
The future of vulnerability research: human expertise amplified by AI
At Searchlight Cyber, we have seen how frontier models can accelerate vulnerability research when they are used by experienced researchers with the right technical context. The models are powerful tools, but their value comes from how they are directed, challenged and validated. They can help researchers explore larger codebases and move through investigative work faster, while human expertise remains essential to understanding what matters and proving that an issue is real.
Our research into WordPress, which led to the discovery of the vulnerability chain known as wp2shell, is a clear example. In July 2026, the team identified an unauthenticated remote code execution chain in WordPress Core, the first vulnerability of its kind in the platform in almost a decade. Given WordPress's enormous footprint, the issue potentially affected several hundred million websites.
Frontier AI models played a role in accelerating the research. The team used GPT-5.6 Sol Ultra during an investigation that took around ten hours and approximately $25 in compute costs. But this was not a case of just asking a model to find a vulnerability and passing on whatever it returned directly to WordPress to deal with. The work relied on researchers' understanding of the software, their ability to direct the investigation and their technical judgment in validating the result.
The outcome had practical defensive value. Searchlight customers received notification and mitigations before a public patch was available. Once the patch was released, proof-of-concept exploits began circulating within hours, followed by mass exploitation in less than two days. The research demonstrated how AI can help defenders identify and mitigate vulnerabilities before they become publicly disclosed or widely exploited.
The value of identifying a vulnerability early depends on whether organizations can establish their exposure and put effective mitigations in place before attackers exploit it.
The future of vulnerability research is not a simple choice between human researchers and AI. Researchers can combine the scale and speed of models with the experience needed to identify meaningful issues, validate their impact and understand how they can be exploited in real environments.
What does this mean for organizations?
For security teams, the implications go beyond keeping up with a growing list of newly discovered vulnerabilities. AI is accelerating several parts of the attack lifecycle, from vulnerability discovery through to exploit development, while many organizations are still working through processes designed for a slower-moving threat landscape.
The exposure window: the period during which an organization remains exposed to a vulnerability before it is remediated, is already a significant challenge to close quickly. Attackers can increasingly weaponize vulnerabilities within hours, while defenders may need to identify affected assets, assess risk, assign ownership, coordinate remediation and verify that the issue has been resolved. AI adds pressure to a process that is already struggling to keep pace.
There are several questions organizations need to be able to answer:
- Do we know what we have? Organizations need continuous visibility of their external attack surface, including assets and services that change, appear or disappear rapidly.
- Do we know what is genuinely exploitable? A vulnerability's severity score does not establish whether it is reachable or exploitable in a particular environment. Security teams need to be able to focus on high-confidence, validated exposures and understand their practical impact.
- Do we understand what attackers are targeting? Vulnerability intelligence needs to be considered alongside real-world threat intelligence, including exploit availability, activity on criminal forums and dark web marketplaces, and the tactics and techniques being used by threat actors.
- Can we act quickly enough? Identifying an exposure is only useful if the organization can prioritize it, get it to the right team, implement a fix or mitigation, and verify that the risk has been reduced.
Knowing that a vulnerability exists is not the same as knowing whether an organization is exposed to it. This is the gap that traditional vulnerability management and attack surface management approaches can struggle to address. Periodic discovery, theoretical severity scores and reactive patching are not sufficient when the threat landscape and an organization's attack surface can change faster than those processes can respond.
AI in Exposure Management
The potential for AI in exposure management goes beyond discovering vulnerabilities at greater scale. Agentic AI can change how organizations investigate and manage exposure, with agents autonomously examining indicators and potential weaknesses across the attack surface, testing them in much the same way a skilled offensive researcher would.
Their effectiveness depends on the quality of the context they can access. They need a continuous and accurate understanding of an organization's attack surface, the capability to validate what is genuinely exploitable, and the ability to generate clear reproduction steps an organization can be confident of. Without that context, autonomy risks producing more noise at greater speed. With it, AI could help organizations move from simply identifying potential weaknesses to continuously understanding and reducing their real exposure. A lot is said about cybersecurity favoring the attacker: "attackers only need to be right once, we need to be right every time". But while an attacker has to discover the target’s environment, defenders (should) already know it. Attackers have AI, but an organization has AI in addition to context.
Done effectively, agents taking on the operational work that follows discovery can radically compress Mean Time to Remediation, reducing the time between identifying a genuine exposure and closing it.
Turning AI's research potential into preemptive security outcomes
Project Glasswing made waves in the industry upon its initial announcement. But its progress also demonstrates that the industry cannot measure success by the volume of vulnerabilities a model can find. The important thing is whether those findings lead to meaningful improvements in security.
The same applies to organizations adopting AI-assisted security tools. Faster discovery will not necessarily reduce risk if teams cannot validate findings, understand their relevance to their own environments and remediate them faster than attackers can act.
The pace of change is rightfully a concern for security leaders. As AI continues to accelerate vulnerability research and exploitation, organizations need to focus on moving beyond reactive approaches to vulnerability management. That means continuously understanding their attack surface, validating which exposures are genuinely exploitable, connecting that knowledge to real-world attacker activity, and prioritizing remediation based on what poses the greatest risk.
Ultimately, the measure of progress is how effectively organizations reduce the time they remain exposed to the vulnerabilities attackers can exploit.
That is the shift towards Preemptive Threat Exposure Management: understanding and reducing real exposure before attackers can take advantage.
To understand what to do next, read our report, AI vs. the Patch Window: What Has Changed, and What Hasn't. Explore how AI is changing the speed, scale and economics of cyberattacks, and what organizations need to do to keep pace.
.jpg)



.avif)


