One of the world’s largest hedge funds now runs its portfolio with a cyber catastrophe in mind. In a research note titled “Current Efforts Aren’t Enough to Slow the AI Cyber Virus,” Bridgewater Associates says AI is helping attackers exploit software bugs faster than defenders can fix them, and that it is “managing money based on a one-in-three chance of an economically destabilizing attack within the next year.” The note, published on November 13 and trending on X this week, comes with charts showing estimated cyber incidents more than doubling since late summer. Here is what Bridgewater found, how it measured it and where the evidence is thinner than the headlines.
Quick Take
- What: Bridgewater researchers Hudson Attar and Arham Habib published a note arguing that AI-enabled cyberattacks are spreading “exponentially with limited controls, like a virus.”
- The numbers: Bridgewater’s estimate of total cyber incidents (60-day average, annualized) climbs from roughly 4,000 a year in mid-2026 to nearly 10,000, with the turn marked at the open-weight release of GLM 5.3.
- Why it matters: Zero-day exploits have roughly doubled, and the average time from public disclosure to exploitation has fallen from about a month to 11 days, according to the note.
- The bet: Bridgewater says it manages money assuming a one-in-three chance of an economically destabilizing attack within a year.
- The catch: The incident series is Bridgewater’s own model built from public reports, and the timing link to one open model is a correlation, not proof of cause.
In plain English
A big investment firm studied computer attacks and says they have risen sharply in the last few months. It believes AI tools now help criminals find and use weaknesses in software much faster than companies can repair them. The firm thinks there is roughly a one in three chance that a single attack does serious damage to the economy within a year, and it is investing with that risk in mind. Its figures are estimates built from public reports, so the exact numbers are uncertain.
What happened
Bridgewater published the note on its Research and Insights site on November 13, 2026, with a downloadable PDF and six charts. The firm says it built “a time series of cyber incidents, a taxonomy for linking incidents to root causes, and a real-time tracking pipeline of exploits.” It scraped public disclosures, media coverage and corporate filings and adjusted them “to account for typical reporting lags for different types of players across jurisdictions.”
Its central warning: “AI cyber risks, cyber-hacking ability using open-source models has reached dangerous levels, and we are managing money based on a one-in-three chance of an economically destabilizing attack within the next year.” Bridgewater adds that it does not think “the potential scale of cyber warfare is well appreciated because of how hard it is to track these attacks systematically.”
The note went viral this week after market accounts summarized the charts:
Cyberattacks are going parabolic, per Bridgewater.
— unusual_whales (@unusual_whales) November 19, 2026 · View on X
Estimated total cyber incidents have more than doubled since late summer, from about 4,000 a year to nearly 10,000 on a 60-day annualized basis. Over most of the past decade they rose slowly, from about 1,500 in 2017 to roughly 4,000 by mid-2026.
What the charts show
Bridgewater’s figures come mostly from its charts, which we read directly from the PDF:
| Bridgewater measure (annualized) | Earlier level | Latest level in the note |
|---|---|---|
| Estimated total cyber incidents (60-day average) | About 1,500 (2017), about 4,000 (mid-2026) | Nearly 10,000 |
| Verified AI-driven cyber incidents | Under 25 a year before 2026 | More than 200 |
| Known vulnerabilities exploited (unique) | About 300 before the GLM 5.3 release | About 450 |
| Zero-day exploits | About 50 | Roughly doubled, peaking above 100 |
| Estimated vulnerabilities found (60-day average) | About 50,000 (2025) | About 180,000 |
| Time from disclosure to exploitation | About a month | 11 days |
The charts mark two model releases as reference points: the Claude Mythos Preview release and the GLM 5.3 open-weight release in late summer. The second chart labels two clusters of verified AI incidents: “Korean bank and card theft AI-agent campaigns” and “OpenAI and Anthropic test agents break out.” Cybersecurity professionals, Bridgewater notes, dubbed the wave of exploits in October “Black October.”
Background: the defender’s window is closing
The note frames the situation as an arms race. Frontier labs have tried to use what Bridgewater calls the “defender’s window,” the period in which closed frontier models are much better at cyber work than freely available open models, to fix critical software first. It names three such programs: Anthropic’s Glasswing, OpenAI’s Daybreak and Google’s Fairwind. These efforts have driven a record number of reported vulnerabilities (we covered VulnCheck’s count of 442 CVEs credited to Glasswing).
The problem, in Bridgewater’s words: defenders “need to fix all of the bugs, not just most of them,” while attackers only need one. It calls the head start between Glasswing’s launch and this summer’s cyber-capable open models “a one-time gift that we’re unlikely to see replicated.” Criminal use of AI tools is already visible elsewhere, for example in the ARTEX attack platform that Check Point found in criminal campaigns.
How the attack cycle works now
Bridgewater describes three mechanisms that tilt the balance toward attackers:
- Patch races: Fixing open-source software “necessarily exposes the underlying vulnerability,” and every downstream user must install the fix before attackers exploit it. AI makes it easier to win that race, which is why the disclosure-to-exploitation window has shrunk to 11 days.
- More code, more bugs: Developers use AI coding tools to add “millions of lines of new code,” so the attack surface grows faster than it can be secured. Bridgewater shows its own merged pull requests rising to an annualized pace of around 70,000, with the steepest rise after Claude Code’s release in February 2025.
- Under-resourced maintainers: Much of the internet runs on open-source software maintained by volunteers and nonprofits with limited resources to find and fix vulnerabilities.
The note also offers one reassurance: threat actors “are not yet broadly deploying fully autonomous attack workflows,” and a human remains in the loop. But it warns that open models “are not far from the capability level where that is likely to change.”
How governments are responding
Bridgewater calls policy responses “directionally helpful but insufficient.” It lists Japan’s new active cyber-defense powers and an emergency package of ¥10 billion (about $63 million), South Korea’s October legislation with a 24-hour incident reporting deadline, mandatory executive-level security chiefs and fines of up to 3% of sales for repeated incidents, the UK’s Cyber Security and Resilience Bill, and a US program for vetted private companies to act against foreign cybercriminal groups under federal oversight. None, it argues, fixes the dependence on thinly resourced open-source projects.
Pros and cons
For users
- Pro: More attention from big investors can speed up funding for security and push companies to patch faster.
- Con: If Bridgewater is right, outages and data leaks at banks, airlines and online services could become more frequent.
For developers
- Pro: AI tools that find bugs are now widely available to defenders, and patch volumes are at record highs.
- Con: Maintainers face a flood of reports and an 11-day average window before a disclosed bug is exploited, and AI-generated code adds new bugs of its own.
For companies
- Pro: The note gives security teams a concrete, finance-world argument for budget and faster patch cycles.
- Con: Firms without the resources for preemptive scanning are the most exposed, and rising incident counts may lift insurance and compliance costs.
What X is saying
Reactions split between alarm and skepticism about the GLM 5.3 link. Security researcher mRr3b00t (@UK_Daniel_Card) questioned the framing:
‘Bridgewater identifies the release of the high-powered GLM 5.3 model as the point at which cyber attacks started jumping. PHOTO: GETTY IMAGES’
— mRr3b00t (@UK_Daniel_Card) November 19, 2026 · View on X
I mean… this doesn’t sound right….
I wonder if I can show something with @Huntio
User @HiiiMason took a middle position: “Timing isn’t proof, but defense had better be speeding up too.” Grok, asked to summarize the note on X, wrote that “Bridgewater tracks a real surge in cyber incidents after GLM-5.3’s open weights released late summer” and that “Offense has the edge: attackers need one win, defenders must patch all.” In the same reply Grok also claimed Anthropic had found GLM 5.3 builds exploits “nearly on par” with Claude Mythos; that claim is not in Bridgewater’s note, and we have not verified it.
Bridgewater tracks a real surge in cyber incidents after GLM-5.3’s open weights released late summer. Anthropic found it builds working exploits nearly on par with their restricted Claude Mythos, with safeguards easy to strip. AI-driven attacks, including Korean bank data thefts via agents, rose from near zero. Offense has the edge: attackers need one win, defenders must patch all. Bridgewater prices a 1-in-3 chance of a major economic hit in the next year.
— Grok (@grok) November 19, 2026 · View on X
Critical analysis
Bridgewater deserves credit for doing something few have tried: building a systematic incident series and saying openly how it trades on it. The supporting signals point the same way: more known exploited vulnerabilities, more zero-days and a shorter window between disclosure and attack.
But readers should keep three caveats in mind. First, the headline incident count is a modeled estimate from public reports, adjusted for reporting lags, and the note does not publish the underlying dataset. Sharp moves at the end of such a series are the most sensitive to those adjustments. Second, marking the GLM 5.3 release on every chart invites a causal reading that the note itself does not prove; many things changed this autumn, including record disclosure volumes from defensive programs. Third, “one-in-three” is a portfolio risk assumption, not a forecast with a published method, and Bridgewater’s own disclaimer says its positions “may, and often will, vary from its conclusions.” The verified AI-driven incident count, around 200 a year, is also small next to nearly 10,000 total incidents, so most of the surge is not directly attributed to AI.
What’s still open
- The follow-up piece Bridgewater promises on where it sees economic impacts.
- Whether independent incident trackers confirm the size and timing of the surge.
- How much of the rise comes from AI-driven attacks versus more reporting and more disclosed bugs.
- Whether frontier labs and governments fund open-source maintainers at the scale the note calls for.
About the analysis
This independent i10x analysis is based on Bridgewater Associates’ research note and its PDF charts, plus public posts on X, as of November 20, 2026. Chart values are read from Bridgewater’s figures and are approximate.
Sources (plain text, no links):
- Bridgewater Associates: “Current Efforts Aren’t Enough to Slow the AI Cyber Virus” by Hudson Attar and Arham Habib (November 13, 2026), web version and PDF
- Posts on X by unusual_whales (@unusual_whales), mRr3b00t (@UK_Daniel_Card), @HiiiMason and Grok (@grok) (November 19, 2026)
i10x Perspective
When a macro hedge fund starts pricing cyberattacks like a recession risk, the conversation has moved from IT departments to boardrooms. Bridgewater’s numbers are estimates, and the open-model link is unproven, but the direction matches what security teams describe: faster exploits, more code and too few people patching it. The most useful takeaway is not the one-in-three figure; it is the 11-day window. Follow the AI Security section for how this develops.



