a16z: AI agents exploited 395 orgs in 26 seconds as exploit costs collapse to hours

a16z Newsletter · rss · 2026-11-05

a16z's long read argues AI has broken the offensive-cyber bottleneck: turning a vulnerability into a working exploit now takes minutes, not months, while patching hasn't sped up.

Key incidents

One operator pointed hundreds of AI agents at a print-management server: the agents built a test lab, wrote exploits for 2 bugs disclosed 4 days earlier, and went from empty workspace to first real break-in in under 4 hours. 11 orgs fell in 26 seconds; 395 organizations across 48 countries were compromised, half of them schools — using a cheap open-weight model.
Two critical PHP bugs leaked pre-patch: attacks hit Amazon's decoy servers 4 days before disclosure, with working shells within 3 days. Amazon's CISO says the 7/30-day patch SLA "assumes a delay that no longer exists."

The data: Mandiant shows disclosure-to-attack time falling from 63 days (2018-19) to 5 days (2023), and attacks now arrive 7 days before disclosure (2025). Only 6% of CVEs were ever exploited in the wild — mostly because attackers lacked time. AI removes that constraint.

Conclusion: no internet apocalypse, but worm-era (2003) and WannaCry-style (2017) outbreaks will hit unpatchable systems — industrial controllers in energy, water, hospitals — and targets once "too small to matter," since attackers no longer need to want a target. Also cited: DIVD breached by an agent escalating "in seconds," Google's first AI-developed criminal zero-day, and NSA/CISA warnings on AI scripts targeting Siemens controllers.

Original post →

a16z: AI agents exploited 395 orgs in 26 seconds as exploit costs collapse to hours

More from AGI Musings