AI finds the flaws faster than anyone can safely fix them
Recent outages were not caused by rogue AI but by rushed patches, while a VMware bypass and a Cisco root flaw are being exploited in the wild.
- AI is finding vulnerabilities faster than organizations can safely fix them, and Mandiant's M-Trends 2026 reports a mean time to exploit of negative seven days — the exploit arrives before the patch does. m:…7676
- Recent outages affecting air traffic control, Microsoft 365, and banking systems were not caused by rogue AI or a coordinated campaign; defenders are skipping best practices like thorough testing and gradual rollout to patch systems quickly. m:…7676
- A VMware vCenter authentication bypass, CVE-2026-123890, is rated CVSS 9.8 Critical, was exploited by the threat actor ARMORY CASTLE in early September 2026, and has no workaround published by Broadcom. m:…8465
- A Cisco Expressway remote code execution flaw, CVE-2026-76466, is rated CVSS 9.8 Critical, had a public proof of concept published on 2026-11-05, was reported exploited in the wild on 2026-11-19, and has no workaround listed by Cisco. m:…8463
The outages were not the AI
Recent global outages affecting air traffic control, Microsoft 365, and banking systems were not caused by rogue AI, but by the pressure AI places on defenders. AI is finding vulnerabilities faster than organizations can safely fix them, and the result is a patching race that the defenders are losing on both ends: too slow to close the hole, too fast to test the fix. m:…7676
Defenders are skipping best practices like thorough testing and gradual rollout to patch systems quickly. The Washington Sun reported that bank executives and the NSA told Congress that rushed patches caused the recent outages. The outages were a human process failure under machine speed, not a machine going rogue. m:…7676
The outages had a concrete shape. A defect in Microsoft 365 gateway servers caused 80% of servers to become unresponsive in November 2026, and a United Airlines system outage in November 2026 resulted in more than 400 delayed departures. Most enterprise software outages originate from unknown connections rather than dramatic mistakes, which is why a routine patch can take down a system nobody knew was connected to it. m:…9573
Why patches are rushed
The clock has collapsed. Mandiant's M-Trends 2026 reports a mean time to exploit of negative seven days, which means exploitation now typically begins before a patch is even available. The FBI issued a FLASH advisory regarding the ARTEX autonomous attack framework, and ARTEX can move from initial targeting to privileged access in under three hours. Palo Alto Networks' Unit 42 found that AI tools target long-standing weak points like exposed APIs and stored credentials. m:…7676
The aftermath is its own attack surface. Attackers are using the recent outages to launch phishing and voice-based social engineering scams, and the FBI recommends inventorying and reducing external attack surfaces to counter autonomous tools. OpenAI models broke into Hugging Face during internal testing, which is a reminder that the same capability is available to both sides. m:…7676
Blitzy, which maps software estates, explains why the rushed patch breaks things. Gartner identifies hidden dependencies and configuration drift as primary drivers of technical debt risk, and most validation testing confirms only the code touched, rather than the system-wide dependencies. AI coding tools often lack the context window to see dependencies beyond the files currently being edited, and Atlassian reports that one-third of AI-generated code is discarded entirely. The fix is applied to code the tester can see; the outage comes from the dependency nobody mapped. m:…9573
Two flaws being exploited now
The first is in the virtualization layer. A flaw in the vCenter Directory Service allows an unauthenticated attacker to recover an SRP session key, and Broadcom rates CVE-2026-123890 as a CVSS v3.1 9.8 Critical vulnerability. The threat actor ARMORY CASTLE exploited CVE-2026-123890 in early September 2026. m:…8465
There is no easy way out of it. Broadcom has not published a workaround for CVE-2026-123890, and the vulnerability is caused by predictable UUIDv1 timestamp generation in the SRP authentication interface. The exploitation was a later stage in an intrusion that began with the compromise of a BeyondTrust Remote Support appliance. Notably, CVE-2026-123890 is not currently listed in the CISA Exploited Vulnerabilities catalog as of 2026-11-20, which means teams relying on that list alone would miss it. m:…8465
The second is in the network edge. A static maintenance identity and cookie in the database service of a clustered Cisco Expressway allows unauthenticated attackers to run commands as root, and Cisco rates CVE-2026-76466 as a CVSS v3.1 9.8 Critical vulnerability. F6 reported in-the-wild exploitation of CVE-2026-76466 on 2026-11-19, two weeks after a public proof of concept was published on 2026-11-05. m:…8463
The details are unglamorous and that is the point. The maintenance identity used in the exploit is named wombat at 127.0.0.1 with a cookie value of wombat, and Cisco lists no workaround for the flaw. Attackers used the Expressway vulnerability to pivot laterally, scanning SMB resources and interacting with SQL Server and WinRM, and the F6 report describes the compromise of four Expressway devices in total. The first fixed release is X15.5.20. m:…8463
The one that was caught in time
Not every hole is being exploited. Chrome 148 fixed two high-severity sandbox escape vulnerabilities, CVE-2026-9997 and CVE-2026-9998. The first is a use-after-free vulnerability in the browser process affecting Windows, Linux, and ChromeOS, and the second is an integer wraparound vulnerability in Skia that affects Android WebView. Neither vulnerability has known exploitation in the wild as of 2026-11-20, and exploitation of these vulnerabilities requires a separate renderer compromise. m:…8464
The contrast is instructive. Google's Project Fortify report for CVE-2026-9998 states that the project's tooling could not produce a functional proof of concept, and Microsoft Edge 148.0.3967.97 is the fixed build according to HKCERT. The difference between the Chrome flaws and the VMware and Cisco flaws is not severity — all of them are serious — it is whether attackers got there first. m:…8464
What defenders should do
The advice from both sides of this week's material points the same way: shrink the surface, and map the dependencies before you patch. The FBI recommends inventorying and reducing external attack surfaces, and Blitzy's answer is to map the entire software estate including files, dependencies, service boundaries, and business rules, with a model that refreshes on every merge to prevent drift. m:…7676 m:…9573
The uncomfortable part is the trade-off nobody has solved. With a mean time to exploit of negative seven days, waiting to patch is dangerous, and with hidden dependencies everywhere, patching fast is dangerous too. The teams that survive that squeeze will be the ones that know what their systems depend on before the patch lands. m:…7676
What is still unresolved
- Whether the VMware flaw reaches the CISA exploited-vulnerabilities catalog and how many unpatched vCenter instances remain exposed
- Whether rushed patching causes more outages than delayed patching over a full year
- Whether autonomous attack frameworks like ARTEX change the economics of patching for smaller organisations
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.