Bridgewater: "Current Efforts Aren't Enough to Slow the AI Cyber Virus"

The world's largest hedge fund says AI-enabled cyber warfare is "spreading exponentially" with limited controls, like a virus, and that it is managing money on a one-in-three chance of an economically destabilising attack within the next year. Published 13 Nov 2026 by Hudson Attar and Arham Habib, and picked up widely on X on 20 Nov. Bridgewater built its own incident time series and exploit-tracking pipeline, so this is original data work, not commentary on other people's numbers.

What Bridgewater says

  • The bet: a one-in-three chance of an economically destabilising cyberattack within a year. This is a portfolio-positioning assumption, not a forecast with a published method.
  • Open-source models are the trigger. Cyber-capable open-weight models (the charts mark the GLM 5.3 weight release) are described as the point where attacks inflected upward. Quartz reports Bridgewater sees a timing coincidence, not proven causation.
  • Attackers vs defenders: defenders must fix every serious bug, attackers need one. Zero-day exploits have roughly doubled since cyber-capable open-source models appeared.
  • Patch window collapsed: the average time from disclosure to exploitation fell from about a month to 11 days.
  • Defence is working but not enough. Programs like Anthropic's Glasswing, OpenAI's Daybreak and Google's Fairwind drove a record surge in found bugs. The report calls the head start before cyber-capable open models arrived "a one-time gift" that will wear off.
  • A human is still in the loop. Attackers are not yet broadly running fully autonomous workflows, but open models are close to the point where that changes.
  • The attack surface is growing: AI coding tools add millions of lines of new, buggy code. Bridgewater's own merged pull requests are charted as an example.
  • Policy is lagging: Japan (¥10bn emergency package, active cyber-defence powers) and South Korea (24-hour reporting, fines, mandatory CISOs) have moved. The US and UK steps are limited. The core gap is volunteer-run open-source software that much of the internet depends on.
  • It names Black October as the start of the current wave and notes it lacked details on the recent Microsoft outages and vishing campaigns against financial firms.

Charts

Estimated total cyber incidents, 60-day moving average annualised, 2017 to late 2026, rising from about 1,500 to nearly 10,000 with a near-vertical jump after the GLM 5.3 weight release Verified AI-enabled cyber incidents per year: intentional AI-enabled attacks reach about 180, total AI-driven incidents about 210, rogue AI agent episodes about 25, all surging after mid-2026. Annotated with Korean bank and card theft AI-agent campaigns and OpenAI and Anthropic test agents breaking out Two charts: estimated cyber vulnerabilities per year rising to about 175,000, and vulnerabilities by severity with the 6-8 and 8-10 bands rising steeply, marked with the Claude Mythos Preview and GLM 5.3 releases Two charts: unique known vulnerabilities exploited per year rising to about 450, and zero-day exploits per year roughly doubling to about 100 after the GLM 5.3 weight release Bridgewater merged pull requests, annualised pace, rising to about 70,000 since Claude Code's release on 24 Feb 2025, driven by new repositories

Charts © Bridgewater Associates, reproduced here as an archive copy with attribution. Bridgewater's report carries a no-redistribution notice, so check before reusing the images elsewhere.