AI systems can now produce working exploits in under an hour for bugs that take weeks or months to patch across deployed systems, according to a new analysis from Andreessen Horowitz published in early 2026. The firm argues that the widening gap between vulnerability discovery and remediation can't be closed by faster patching alone, and that the cybersecurity products and policies that matter over the next five years will be those protecting systems even when rapid patches aren't possible. The analysis builds on earlier work showing that turning a bug into a functioning exploit, once a bottleneck requiring skilled human labor, became cheap in 2025 and 2026 while patching speeds barely moved.
The report cites median full remediation time at 43 days according to Verizon's 2026 breach report, with only 26 percent of known-exploited vulnerabilities fully remediated in the dataset. Average flaw-remediation time reached 252 days in Veracode's 2025 report, up from 171 days five years earlier. In October, an unreleased AI model produced a remote, no-login exploit for a Linux kernel bug in under an hour, targeting a vulnerability that had been fixed upstream weeks earlier but still lacked a patched package in two widely used server distributions. That same month, two critical PHP bugs leaked before their fix, with exploit attempts starting four days before public disclosure and leaving between 2.6 and 4.4 million PHP installations exposed, most on versions that will never receive a fix. Industrial control systems face even longer cycles: U.S. grid rules allow 35 days to evaluate a patch and 35 more to apply it, while Dragos found that about a quarter of 2025's industrial advisories shipped with no patch or mitigation at all. Medical devices compound the problem further, with one 2025 survey finding 89 percent of hospital systems had connected devices with a known-exploited vulnerability, many running on unsupported operating systems where patches can trigger re-certification requirements.
The authors argue that for equipment that can't be patched quickly, protection must happen off the device. According to the analysis, when researchers at Censys found nearly 400 water-utility control screens exposed online, including 40 allowing control without login, working with the EPA and vendor secured a quarter within nine days and fewer than 6 percent remained exposed by May 2025, with no patch involved. The report states that in one study of infected consumer devices, a European internet provider's quarantine-and-notification process cleared 92 percent of infections within 14 days, while email warnings did no better than saying nothing. Backups have become a primary target themselves: in Sophos's 2024 survey of ransomware victims, 94 percent said attackers tried to compromise their backups, and 57 percent of those attempts succeeded, with victims whose backups were compromised reporting median recovery costs of about $3 million against $375,000 where backups survived.
The firm identifies four defensive approaches that work when patching can't keep pace: removing exposure from the open internet, segmenting what must stay connected, blocking attacks in the network at machine speed through virtual patching, and eliminating entire classes of bugs through memory-safe languages and formal verification. Google's shift to memory-safe code reduced memory-safety bugs from 76 percent of Android's vulnerabilities in 2019 to 24 percent in 2024, the report notes. The analysis describes a future where live fixes are applied to running systems within hours through edge rules, in-process hooks, or memory patches, with AI coding agents writing version-specific mitigations at machine speed while also producing permanent code fixes on a parallel track. For industrial systems, medical devices, and other long-tail infrastructure, the economics push toward sheltering under a platform or removing exposed infrastructure from the internet entirely, since defending a single asset costs about what it always did while attacking keeps getting cheaper. The firm sees six product categories positioned to win: systems that fix vulnerabilities live then fix the code, detection without rules that responds first, per-request identity judgment instead of per-login, backup posture management that proves recovery rather than assumes it, unified software supply-chain controls at the point of install, and long-tail platforms combining exposure management with secure remote access for organizations that will never run security teams. The last worm era ended when one company took ownership and fixed it, the authors note, but nobody owns this one.
The shift from human-gated patching to AI-assisted live mitigation represents a fundamental reordering of defensive priorities, where proving a system can be restored may matter more than preventing every intrusion. Organizations betting their resilience on patch speed alone are likely mispricing the risk that their infrastructure will be exploited faster than it can be secured.