AI agents make security through obscurity obsolete, researchers say

The Register reports that "security through obscurity," the old assumption that a system stays safe as long as its design and flaws remain hidden, no longer holds now that AI agents can analyze almost any codebase or protocol quickly. Brett Leatherman, assistant director of the FBI's Cyber Division, said open source libraries that run on 80 percent of web servers and had been stress tested by the community for a decade were assumed secure, until the latest AI models found significant vulnerabilities in them. Trend Micro Zero Day Initiative chief bug hunter Dustin Childs made a similar point the day after Microsoft's Patch Tuesday addressed 974 CVEs, noting that many of the patched components, such as the Telnet client, the Windows RNDIS USB networking protocol, 1980s era NFS Portmapper, and the Vista era Link Layer Topology Discovery protocol, had not been discussed in years before AI-assisted bug hunting surfaced them.

The same shift helps attackers. The article describes at least four espionage crews, most suspected of links to China, exploiting Chromium's "patch gap," the window between an upstream fix and the downstream stable release, with an AI-developed exploit kit built shortly after maintainers published the patch. Former US National Cyber Director Chris Inglis and John Hultquist, chief analyst at Google Threat Intelligence Group, told The Register at Black Hat in August that they are worried about the effect on operational technology and industrial control systems, which historically relied on obscure protocols and proprietary hardware to stay safe from attackers who lacked specialist knowledge. Hultquist said AI removes that barrier because it is good at technical troubleshooting and at explaining obscure systems, and that the expertise protecting these systems has long lived in a small number of people's heads. A couple of weeks after Black Hat, five US agencies said attackers used AI-generated exploitation scripts to break into internet-exposed Siemens S7 Series programmable logic controllers at water, manufacturing, energy, and other critical facilities, warning that "this is not a theoretical risk, it is an active threat."

Katie Moussouris, founder and CEO of bug bounty consultancy Luta Security, said security through obscurity was never a winning strategy because a determined attacker who turns their attention to an organization will eventually find what is there to find. She added that AI now lets people probe a tech stack they are not personally familiar with, since the AI has effectively ingested everything. But she argued the real bottleneck was never finding bugs, it is triaging, prioritizing, and actually fixing them, and that AI has not caught up on the defensive side.

Two recent studies back that up. 1Password's research team took six CVEs disclosed since March and generated 6,080 patches for them using OpenAI's ChatGPT-5.5 and Anthropic's Opus 4.8. Director of Security Research Keith Hoodlet reported an average success rate of just 26.0 percent for a patch that fully resolved the vulnerability without materially changing application behavior, and said even patches that did fix the flaw broke the application's behavior, for example by flipping an "allow list" into a "deny list," 20 percent of the time; overall, 53.9 percent of the generated patches either failed to fix the bug, introduced a new vulnerability, or both. Separately, application security firm Veracode found an average security pass rate of just 56 percent for AI-generated code across more than 100 models and 80 coding tasks. Moussouris said organizations that only throw more resources at finding and fixing individual bugs, rather than looking at the process failures behind recurring flaw types, such as fixing injection flaws at the template level earlier in the CI/CD pipeline, will "die on the treadmill."

Key facts

Microsoft's latest Patch Tuesday addressed 974 CVEs, including fixes for long-ignored components like the Telnet client, Windows RNDIS, 1980s era NFS Portmapper, and the Vista era Link Layer Topology Discovery protocol.
FBI Cyber Division assistant director Brett Leatherman says AI models found significant vulnerabilities in open source libraries that run on 80 percent of web servers and had been considered secure for a decade.
At least four espionage crews, most suspected of links to China, used an AI-built exploit kit to hit Chromium's patch-gap window between the upstream fix and the stable release.
Five US agencies warned that attackers used AI-generated exploitation scripts to break into internet-exposed Siemens S7 Series PLCs at water, manufacturing, and energy facilities, calling it an active threat rather than a theoretical one.
1Password's study of six CVEs found patches from ChatGPT-5.5 and Opus 4.8 fully fixed the bug without side effects only 26.0 percent of the time, and 53.9 percent either failed to fix it or introduced a new flaw; Veracode's separate study of 100+ models found a 56 percent security pass rate.

Why it matters

The article argues that the old defensive crutch of hiding a system's design or relying on its obscurity is gone, because AI agents can now analyze almost any code, protocol, or device quickly, whether the analyst is a defender or an attacker. That is producing a genuine surge in disclosures and patches, not just anecdotal reports: a single Patch Tuesday fixing 974 CVEs, some in components untouched for decades, is offered as direct evidence.

Who it affects

Open source maintainers face a bigger disclosure backlog as AI-assisted researchers surface long-buried bugs. Major vendors like Microsoft and Adobe are shipping unusually large patch batches. Critical infrastructure operators in water, energy, and manufacturing are named specifically, since their industrial control systems and PLCs, like the Siemens S7 series, historically depended on obscure, proprietary tech that is no longer a meaningful barrier. Bug bounty programs and security teams doing triage are also affected, since Katie Moussouris argues the bottleneck has shifted further toward fixing and away from finding.

How to use it

The practical guidance in the piece comes from Moussouris: rather than only counting bugs found and fixed, organizations should look for patterns across vulnerability types, for instance treating recurring injection flaws with safer templates earlier in the CI/CD pipeline, to prevent whole classes of bugs at scale. The two cited studies also caution against trusting AI-generated patches without review, since both 1Password and Veracode measured failure or side-effect rates above 40 percent.

How solid is it

The claims rest on named, on-record sources: an FBI Cyber Division assistant director, Trend Micro's Zero Day Initiative chief bug hunter, a former US National Cyber Director, Google Threat Intelligence Group's chief analyst, and Luta Security's founder, plus two specific studies with stated sample sizes (1Password's six CVEs and 6,080 generated patches; Veracode's 100+ models and 80 coding tasks). The source does not give a year for Black Hat, the Patch Tuesday, the Siemens warning, or the two studies, only relative markers like "last week" and "a couple of weeks after Black Hat," and it does not name the four espionage crews, the six CVEs 1Password tested, or the specific models Veracode covered.

Risks and caveats

The piece describes a widening gap: AI is speeding up both bug discovery and exploitation, including against industrial control systems tied to water, energy, and manufacturing, while the same AI tools remain unreliable at fixing what they find, failing to resolve or actively introducing new vulnerabilities in the majority of cases in the 1Password study. The source itself does not identify the specific CVEs, vendors, or espionage crews involved, so those details cannot be independently checked from this article alone.

“The latest models were able to break those and say, 'yeah, there's significant vulnerabilities in here.'”

— Brett Leatherman, assistant director of the FBI's Cyber Division