JPMorgan Says AI Has Raised Cyber Risk Tenfold. What Does That Mean?

Server rack with network cables beside a laptop displaying codeServer infrastructure illustrates the cybersecurity risks discussed in the article.
Jamie Dimon has put a striking number on the cybersecurity pressure created by powerful AI. In an October 6 Bloomberg interview, the JPMorgan Chase chief executive described the risk as having increased tenfold following the arrival of Anthropic's Mythos.

That deserves attention, particularly from the head of a major bank. It also needs careful reading: the tenfold figure is Dimon's assessment of the threat, not a published measurement showing that successful attacks across the banking industry have multiplied by ten.

Finding weaknesses is becoming easier

The concern is that advanced models can help locate software weaknesses and work through the steps needed to exploit them. The same capabilities can help defenders examine systems before an attacker reaches them.

Anthropic built Project Glasswing around that tension. Announced in April, it gave selected organizations access to Claude Mythos Preview to investigate and secure important software. JPMorgan Chase was among the founding participants, alongside major technology and infrastructure companies.

The program's examples included longstanding flaws in widely used software. That distinction matters: AI can expose weaknesses that were already present. Discovering a bug does not mean the model created the underlying defect.

The numbers behind the concern

In its October update to the Cyber Verification Program, Anthropic reported that Glasswing partners identified at least 129,000 verified software vulnerabilities between April and July 2026. Its own open-source scanning found another 5,500 between April and October.

More than 33,000 of those verified findings had been rated high or critical severity. These are figures reported by Anthropic from its program, rather than an independent count of all vulnerabilities or attacks worldwide.

The company also explains that the results draw on partial partner reporting, with different approaches to triage and incomplete information about patches. The figures demonstrate the volume of findings being handled; they do not tell us that every issue has already been fixed.

A longer bug list is only the beginning

For a security team, discovery starts a chain of work. Someone must confirm the affected system, judge the impact, assign responsibility, prepare a fix, test it and get it deployed. A rapidly expanding list can overwhelm that process even when the discovery tool is doing useful work.

The practical question is how quickly an organization can close the gap between a verified finding and a working repair. Counting discoveries alone misses the period during which the weakness remains exposed.

Access is part of the response

Anthropic is now bringing its security programs into a broader verified-access structure, with different permissions and controls for defensive work, authorized testing and especially sensitive systems. Powerful cyber capabilities are being distributed under conditions rather than treated as an ordinary unrestricted consumer feature.

That is also why the distinction explored in The Next AI Model Is Here. Can I Actually Use It? matters. Announcing a capability and making it broadly available are separate decisions.

Dimon's warning captures the urgency. The more useful measure of progress will be whether defenders can turn AI's faster discoveries into repairs before those same weaknesses are used against them.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs