AI is speeding up both sides of vulnerability management: systems can find flaws and draft fixes faster, while attackers can use the same tools to study those fixes and search for ways around them. The practical pressure on security teams is shifting from finding a vulnerability to verifying that a patch closes every path to it.
That tension surfaced in work on vm2, an open-source JavaScript sandboxing library with about 1 million downloads per week. This summer, maintainers released fixes for 34 sandbox escape vulnerabilities, with Anthropic’s Claude contributing to 19 patches. Within weeks, a security researcher pointed an AI agent at one of the advisories and asked it to find a way around the fix. It did.
For Shayaun Nejad, lead product security engineer at Rubrik, the case illustrates why patches need their own security review. While preparing a talk for the [un]prompted AI security conference in San Francisco, Nejad tested vm2’s patch for variants: related flaws that share the same underlying cause. His team had seen bugs return in later testing, sometimes in the same code and sometimes in nearby code, after earlier fixes were marked complete.
“Sometimes the [vulnerability] variant wouldn’t be caught until the model variation got better or the model used chaining analysis at a higher degree once it improved in order to find those variants,” Nejad said. Variant analysis, he added, is becoming “a context magnet” because a disclosed bug gives an AI system useful clues about where and how to look.
A patch can disclose the shape of a flaw
Security fixes reveal information. A code change can show which input was dangerous, which function was affected and what defenders believed the root cause to be. That knowledge can help maintainers search for related bugs, but it can also give attackers a starting point for testing whether the repair is incomplete.
Evidence from patch studies suggests that this is a practical problem, not just a theoretical risk. Researchers at 1Password’s Off-by-1 Labs asked AI agents to generate fixes for six known vulnerabilities in complex open-source projects. About half of the 6,080 candidate patches left at least one attack path open. In some cases, agents missed another instance of the same bug elsewhere in the code.
A separate benchmark, P2Patch, set one AI agent to exploit 101 real vulnerabilities and another to patch them. After the second agent’s fixes, the first tried again. It found a working route in 41 cases, and sometimes succeeded again after further patches.
Specific patch failures show why checking only whether the reported exploit still works can be inadequate. At the [un]prompted conference, OpenAI security researcher Tim de Waal described an AI agent repairing two vulnerabilities in network access point firmware. One replacement function omitted stack canaries, removing a protective measure in the original code. In another fix, the agent filtered double quotes but failed to account for single quotes, leaving a way to bypass the input check.
Human fixes also fail. Trail of Bits reviewed patches submitted for 2,265 vulnerabilities found during its consulting assessments and found that one in eight did not fully resolve the reported issue. In one documented case, a human maintainer and an AI agent working through the Patch the Planet initiative independently wrote fixes for the same memory-safety bug in the freenginx web server. Both introduced a crash during request cleanup.
Those comparisons don’t make AI patching pointless. They show that patch quality needs to be judged by testing the underlying cause and its variants, rather than by whether one known demonstration stops working.
Discovery is outpacing review
The volume of reported vulnerabilities is increasing, putting maintainers under pressure to triage findings, confirm which are real and produce fixes. Anthropic’s Mythos-powered Glasswing program had logged 37,966 findings seven months after launch. Of 7,349 flaws made public, only 217 had been fixed, according to an analysis by vulnerability intelligence firm VulnCheck. Anthropic reported a 94.2% true positive rate, though not every finding is necessarily a security vulnerability.
Anthropic has described the bottleneck as a shift: software security was once constrained by the speed of vulnerability discovery; now, it says, the constraint is how quickly teams can verify, disclose and patch the findings AI systems produce.
The strain reaches large organizations as well as open-source maintainers. Bank of America added 1,000 outside contractors in the second quarter to support patching work, alongside roughly 3,000 existing staff. Chief Executive Brian Moynihan said in October that the contractors were still working on the effort.
At Meta, AI was writing fixes for 40% of critical vulnerabilities by July, and nearly half of those patches went into production unchanged. Yet the time systems remained exposed did not fall, according to Meta software engineer David Molnar. Human review remained necessary as the volume of new code grew by more than 100% year over year; more than 80% of that code was attributed to AI coding agents.
Molnar described reviewers as overwhelmed and said Meta was looking for ways to automate more of the process, including review. Automating the steps that produce fixes can increase the queue that people must inspect, he argued, so teams need to consider the full pipeline rather than accelerate one stage in isolation.
The numbers suggest why that matters. In June, the Forum of Incident Response and Security Teams projected about 66,000 CVE assignments for the year. With two months remaining, the count was nearing 100,000. Microsoft’s October Patch Tuesday addressed 1,415 vulnerabilities, compared with 180 in October 2025.
Closed-source software has the same exposure
Source code isn’t required for AI-assisted patch analysis. Researchers can compare vulnerable and fixed binaries, use decompilation output and identify what changed. That approach has long been part of security research; AI is reducing the time and expertise needed to apply it.
In June, Anthropic researchers tested how quickly an attacker could exploit information released with a patch. Using public material available on patch day, a model produced a working exploit within an hour of a Firefox fix being released. In tests involving Windows kernel vulnerabilities, the researchers supplied vulnerable and patched binaries alongside decompilation output. The approach built eight full privilege-escalation chains from 21 vulnerabilities, at a cost of roughly $2,000 per chain.
A team at Fudan University studied 76 recent patches across 62 closed-source applications from 25 vendors. Its LLM-assisted binary diffing system flagged 63 patches as potentially incomplete. Researchers manually confirmed 56, then found that 26 could be bypassed into new vulnerabilities. The work led to 19 CVE assignments and $28,100 in bug bounties.
Microsoft vice president of security research Taesoo Kim has warned that AI has shortened the time required for binary-diff analysis and exploit development. Some frontier models refuse this kind of work under their guardrails, he said, but open-weight models such as Qwen, Kimi and GLM remain capable alternatives.
That speed also affects patch gaps: the period between public disclosure or a fix appearing in one project and its arrival in related products. The Microsoft Edge Security team has described attackers using information from Chromium fixes before those changes reach downstream browsers such as Edge and Chrome. AI-assisted analysis can compress work that once took weeks into hours, leaving less time for downstream teams to ship updates.
Researchers from Epoch AI used GLM 5.3 to target Firefox releases during such a gap. With about $1,000 in API credits and as little as 12 hours per target, they compromised every release shipped between Sept. 17 and Oct. 26. Mozilla responded by keeping regression tests for security fixes confidential until the fix has shipped. Firefox CTO Bobby Holley said GLM 5.3 was much less effective at exploiting the gap when those tests were unavailable.
Make variant hunting part of the fix
Withholding patch details isn’t a durable answer for projects that depend on public security information. The more direct defense is to make fixes harder to bypass and to test them against the bug’s cause, including related code paths.
Rubrik added automated variant hunting after its team saw supposedly fixed bugs reappear. It now routes simpler issues to AI patching and reserves complex flaws, such as authentication bypasses, for human engineers. Nejad’s point is that defenders have the same context as attackers and can use it first: teams can search for variants as soon as they understand a vulnerability.
OpenAI has also said its Linux vulnerability research and patching process required iterative testing and variant analysis because initial fixes did not always examine related code. Its researchers recommend using models to review human and AI-generated patches adversarially, probing for bypasses, missing attack paths and variants.
The Open Source Security Foundation’s guidance recommends having a separate discovery agent test a patch from an attacker’s perspective. Trail of Bits has released an agent skill that requires testing at least one distinct variant tied to the vulnerability’s root cause before a patch can be submitted.
Terra Security reported an eightfold increase this year in incomplete or broken fixes across open-source projects. Its researchers warn that a patch aimed only at the reported input can create false confidence while leaving the underlying defect in place. A nearby code path, different encoding or alternate input may be enough to expose it again.
The operational takeaway is clear: AI can help teams reduce the backlog, but a generated patch should start a verification cycle, not end one. Teams that pair faster remediation with root-cause analysis, codebase-wide variant searches and adversarial retesting will be better placed to keep disclosure useful for defenders without handing attackers an easy map.