AI cyberattack capabilities double every 4 to 5 months, having caught up with the pace of programming evolution, and the defender's "patch window" is heading toward zero.

a16z partner Matt Bornstein published a 23-year attack-defense dataset showing that large language models now double their autonomous exploit-writing capability every 4-5 months, matching the pace of code generation and collapsing the patch window that defenders have relied on. Mandiant data cited in the report shows the average gap between vulnerability disclosure and first in-the-wild attack fell from 63 days in 2018-2019 to 5 days in 2023, and by 2025 attacks were arriving an average of 7 days before public disclosure. The report cites an August incident in which a single operator deployed hundreds of AI agents against a widely used print-management server, going from an empty workspace to compromising a real victim in under 4 hours and hitting 395 organizations across 48 countries, half of them schools, using a cheap open-weight model rather than a frontier flagship. A second case saw Amazon honeypots attacked 4 days before a PHP vulnerability was publicly disclosed and 12 hours before the official patch, with a working interactive shell produced within 3 days. The report argues the marginal cost of attack is approaching zero, ending the "too small to target" era for millions of SMBs, local hospitals and water systems, while critical infrastructure with physical isolation and redundancy should avoid instant collapse. What to watch next is whether defenders can compress patch cycles and adopt AI-driven remediation before broad worm-style outbreaks, akin to 2003 and WannaCry, hit unpatched legacy systems.