Anthropic’s Mythos reveals a growing security gap: AI finds flaws far faster than companies can patch them

As my colleague Beatrice Nolan has reported, Anthropic’s new AI model, Mythos, has caused a stir among cybersecurity experts and policymakers by saying its new model is so skilled at finding and exploiting software vulnerabilities that it’s too dangerous to release. Instead, it is limiting access to a small group of major technology companies whose software is the foundation for many other digital services, hoping to give defenders time to strengthen their systems.

But this announcement also revealed a growing concern for those defenders: AI is finding flaws far faster than companies could ever hope to patch them. According to Anthropic, Mythos has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. 

Flaws are coming faster than companies can fix

“Vulnerability discovery is outpacing patching,” Shane Fry, CTO and RunSafe Security, told me by email yesterday. AI is accelerating exploit discovery beyond what organizations, especially in operational technology environments (think manufacturing, building systems, industrial control systems and power grids) can realistically remediate, he said. 

Critically, Anthropic says over 99% of what they found has not yet been patched.

“Organizations are already struggling to keep up with patching across both IT and OT environments, and AI is only accelerating that gap,” Fry said. “As vulnerability discovery and exploit development move faster, the idea that you can remediate everything in time just doesn’t hold. The focus has to shift to protections built into the software itself that prevent vulnerabilities from being reliably exploited.”

In a way, an AI tool like Mythos that can find thousands of cybersecurity vulnerabilities a minute is really an “incredibly expensive alarm,” said Tal Kollender, a former hacker and founder of cybersecurity platform Remedio. Finding risk faster than you can fix it, she told me, does not make companies more secure. 

“Don’t get me wrong, what Anthropic is doing is amazing, it’s a wow,” she said, pointing to its game-changing speed of detection. But now companies also need an equal game-changer as far as remediation – that is, fixing the problem once it has been discovered. That’s because fixing vulnerabilities is still a slow, manual process. Teams have to file tickets, patch different systems one by one, track how everything is connected, and make sure they don’t accidentally disrupt the business.

AI is needed to make security fixes faster

Kollender said she began immediately hearing from clients after the Mythos news was released. “They were panicking,” she said. The likely path forward, she explained, is AI-driven systems that don’t just find vulnerabilities, but prioritize, fix, and validate them automatically.

For now, though, the uncomfortable reality is that AI is making it far easier to find weaknesses than to fix them. At least for the next year, Kollender warned, defenders are finding themselves in a race they’re not yet equipped to win. 

With that, here’s more AI news.

Sharon Goldman
sharon.goldman@fortune.com
@sharongoldman

FORTUNE ON AI

Anthropic is facing a wave of user backlash over reports of performance issues with its Claude AI chatbot – By Beatrice Nolan

American Express releases tools to build AI payments—and pledges to pay the price if agents go awry – by Jack Kubinec

After growing up on a dairy farm, this Peter Thiel–backed founder is using AI to save cattle ranching – by Jake Angelo

AI IN THE NEWS

Suspect in attack at Sam Altman’s house aimed to kill OpenAI CEO, warned of humanity’s extinction from AI. CNBC reported that the man accused of throwing a Molotov cocktail at the home of Sam Altman last week has been charged with attempted murder, with prosecutors alleging the attack was planned and motivated by hostility toward AI development. The suspect, Daniel Moreno-Gama, also faces federal explosives and weapons charges after authorities said he targeted Altman’s residence and carried a document outlining his intent to kill the CEO and warning of humanity’s “impending extinction” from AI. Investigators say the document referenced additional AI executives and framed the attack as part of a broader ideological stance against the technology. 

Anthropic hires Trump-linked lobbying firm Ballard Partners. According to Bloomberg, Anthropic has hired the politically connected lobbying firm Ballard Partners amid the company's dispute with the U.S. Department of Defense, which recently designated Anthropic a supply chain risk—a label typically reserved for foreign adversaries. The move came just days after the designation, as Anthropic also mounted a legal challenge and sought to push back on Pentagon demands for broad access to its AI tools. Negotiations reportedly broke down over the company’s insistence on limits, including preventing use of its systems for autonomous weapons or mass surveillance, highlighting growing tensions between AI firms and national security agencies over control and use of advanced models.

The AI revolution in math has arrived. This is a great feature from Quanta Magazine, which argues that AI has crossed a real threshold in mathematics, shifting from a curiosity to a genuinely useful research partner. Systems are now solving competition-level problems, contributing incremental advances on open questions, and helping mathematicians explore large solution spaces far faster than before. High-profile mathematicians like Terence Tao say the field may soon “look and feel” fundamentally different, even as others warn of tradeoffs, including overreliance on AI and a potential loss of deep human intuition. 

EYE ON AI RESEARCH

The striking disconnect between optimistic AI experts and a skeptical public. The 2026 AI Index from the Stanford Institute for Human-Centered Artificial Intelligence highlights a widening gap: AI capabilities and adoption are accelerating rapidly, while governance, infrastructure, and public trust struggle to keep up. Industry now produces more than 90% of leading models, many reaching PhD-level performance, and adoption is widespread, with over half the global population using generative AI and nearly 90% of organizations deploying it. But the U.S. no longer holds a clear lead in performance, with China at parity, even as American firms dominate investment—an advantage that may erode as global talent inflows decline. At the same time, the costs of AI are rising, from energy and water demands to increasing safety incidents and inconsistent safety standards. Perhaps most striking is the growing disconnect: experts remain optimistic about AI’s impact, while the public is far more skeptical, especially around jobs and trust—suggesting the technology is scaling faster than the systems meant to support it.

AI CALENDAR

June 8-10: Fortune Brainstorm Tech, Aspen, Colorado. Apply to attend here.

June 17-20: VivaTech, Paris.

July 6-11: International Conference on Machine Learning (ICML), Seoul, South Korea.

July 7-10: AI for Good Summit, Geneva, Switzerland.

BRAIN FOOD

Could AI transform quantum? Nvidia’s new “Ising” models, released today, are an early glimpse of how AI could help make quantum computing actually useful. Today’s quantum computers are extremely fragile—constantly making errors and requiring painstaking manual tuning—so much so that they still struggle to solve real-world problems. Nvidia’s idea is to use AI to automate those bottlenecks: its open-source Ising models can help calibrate quantum processors and correct errors in real time, tasks that currently take huge amounts of compute and human effort. The bigger picture is that quantum computing may not arrive as a standalone breakthrough, but as a hybrid system—where classical AI systems help stabilize and scale quantum machines. In other words, before quantum transforms AI, AI may first be what makes quantum work at all.