Summary
- AI agents are finding decades-old vulnerabilities in widely deployed software and obscure protocols at record pace, contributing to a 974-CVE Microsoft Patch Tuesday and a massive backlog for maintainers.
- Attackers are using AI to reverse-engineer patches and develop exploits within hours of upstream fixes, closing the patch-gap window that defenders once relied upon.
- OT and ICS environments — historically protected by proprietary obscurity — are now accessible to attackers without specialist knowledge, with US agencies confirming AI-generated scripts were used against internet-exposed Siemens PLCs.
- AI-generated patches fail to fully resolve vulnerabilities an average of 54 percent of the time, according to research from 1Password, and the broader AI code security pass rate sits at just 56 percent across a Veracode study.
- Security experts argue the real problem is not finding bugs faster but fixing broken development processes that generate them at scale.
The assumption of hidden safety is gone
Security through obscurity was never a sound foundation, but for many organisations it served as a quiet supplement to genuine controls — the idea being that if attackers did not know what you were running, they could not easily exploit it. That assumption has collapsed. AI agents are now systematically scanning both commercial and open source codebases, surfacing vulnerabilities that have sat undetected for years or even decades. Brett Leatherman, assistant director of the FBI’s Cyber Division, put it plainly to The Register: libraries that have been stress-tested by the community for ten years and considered robust are yielding significant vulnerabilities when the latest models are pointed at them.
Record patch volumes are an early signal
The practical consequence is already visible in vulnerability disclosure numbers. Microsoft’s most recent Patch Tuesday addressed 974 CVEs — a record. Dustin Childs, chief bug hunter at Trend Micro’s Zero Day Initiative, noted to The Register that components nobody has discussed in years were among those patched: the Telnet client, Windows RNDIS, NFS Portmapper dating to 1980s Unix, and Link Layer Topology Discovery from the Vista era. These are components that persisted quietly in production environments precisely because no one was looking at them. AI is looking now.
Attackers are closing the patch gap
The threat is not confined to defenders using AI to find bugs. Attackers are applying the same tools to reverse-engineer fixes. In one documented case, at least four espionage groups — most suspected of links to China — developed an exploit kit targeting open source Chromium shortly after an upstream patch was released, but before the stable downstream release reached end users. That window, once measured in weeks or months, is now measured in hours. Organisations that factor patch-gap timing into their risk calculations need to revisit those assumptions.
OT environments are the sharpest edge of this risk
The implications for operational technology are particularly serious. ICS and OT systems have long relied on proprietary protocols, specialised hardware, and a narrow pool of expert knowledge as de facto security controls. AI dissolves all three advantages simultaneously. John Hultquist, chief analyst at Google Threat Intelligence Group, told The Register that AI is excellent at technical troubleshooting and navigating obscure systems — precisely the characteristics that have historically kept OT environments off limits to generalist attackers. This is no longer theoretical. Five US agencies confirmed that attackers used AI-generated exploitation scripts to compromise internet-exposed Siemens S7 Series PLCs at water, manufacturing, and energy facilities. The agencies described it as an active threat.
Remediation is the real bottleneck
Katie Moussouris, founder of Luta Security, argues that finding vulnerabilities has never been the hard part. Triage, prioritisation, and remediation are where organisations consistently fall short — and AI is not yet helping on that side of the ledger. Research from 1Password’s security team, testing OpenAI’s ChatGPT-5.5 and Anthropic’s Opus 4.8 across six CVEs and 6,080 generated patches, found an average success rate of just 26 percent for patches that fully resolved the vulnerability without altering application behaviour. In 54 percent of cases, the AI-generated patch either failed to fix the flaw, introduced a new one, or both. A separate Veracode study, covering more than 100 models and 80 coding tasks, found the average security pass rate for AI-generated code was 56 percent.
Patching faster is not enough
Moussouris warns that organisations treating this as a resourcing problem — simply throwing more people at bug discovery and patching — are likely to exhaust themselves without improving their underlying position. The more durable approach is identifying patterns in the types of vulnerabilities being discovered and addressing the process failures that produce them. As an example, she points to injection flaws: rather than patching each instance individually, organisations can address them at scale through safer templates earlier in the CI/CD pipeline. Measuring progress by counting bugs fixed and speed of remediation tells only part of the story. Understanding the categories and root causes tells you what to fix permanently.
Why it matters
CISOs who have tolerated residual reliance on obscurity — whether through undocumented legacy systems, proprietary OT protocols, or simply the low profile of ageing components — are now operating without that buffer. The combination of AI-accelerated discovery and a compressed patch-gap window means exposure periods are shrinking on the attacker side while remediation remains slow and error-prone on the defender side. The OT risk is acute: organisations running internet-adjacent industrial systems need to treat AI-assisted attacks on those environments as a present operational risk, not a future scenario.
What to do now
- Audit legacy and OT systems for internet exposure immediately — components running obscure protocols or proprietary firmware are no longer protected by their low profile.
- Reassess patch-gap assumptions for critical software dependencies, particularly open source projects where upstream and downstream release timing creates exploitable windows.
- Do not rely on AI-generated patches without human validation; current models fail to fully resolve vulnerabilities in the majority of tested cases and can introduce new flaws.
- Shift vulnerability metrics beyond count and speed of fixes to include vulnerability category analysis, which enables process-level interventions in the CI/CD pipeline.
- Treat AI-assisted OT attacks as an active threat category and review whether Siemens S7 or similar PLCs are internet-exposed in your environment.
