Mass cybersecurity disruptions, such as this week’s scramble by governments and companies to patch two critical software flaws, are about to “get a lot worse before they get better,” a top executive at Anthropic PBC has warned.

Banks, airlines and governments around the world pulled down their websites on a precautionary basis this week to prevent them from being hacked after cybersecurity researchers, with the aid of AI, discovered vulnerabilities in the open-source programming language PHP that underpins huge swaths of the internet.

“As bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better,” Jason D. Clinton, deputy chief information security officer for Anthropic, wrote in a post on X on Thursday. “Security leaders, never let a crisis go to waste: please go to your boards and ask for help right now. Not for AI, for the basics.”

AI models are increasingly able to uncover serious cybersecurity vulnerabilities — a capability that helped researchers find the two PHP flaws that led to the proactive service suspensions this week, but could also turbocharge hackers’ ability to exploit weaknesses. That has set a blistering new pace for cybersecurity, collapsing the time that organizations have to respond to AI-discovered threats to close to zero.

So far, there is little evidence that the rise in disclosed vulnerabilities has led to more cyberattacks. Many flaws have been found by technology companies themselves: Google reported 401 of the 433 vulnerabilities identified in Chrome in July. The US government’s Known Exploited Vulnerabilities catalog also shows no meaningful increase in vulnerabilities known to have been exploited.

As yet, there is no public evidence of major successful breaches via the two PHP flaws, but the episode has left a huge backlog of vulnerable systems that security teams must secure. That backlog could snowball as AI tools help uncover fresh vulnerabilities faster than organizations can patch them.

“AI has industrialized discovery, but fixing remains a human bottleneck,” said Chris Hughes, vice president of security strategy at Noma.

Cybersecurity firm Cloudflare Inc. is “seeing millions of attempted attacks using this flaw now across all of our network every single minute of every single day,” Chief Executive Officer Matthew Prince said Thursday in an interview with Bloomberg TV.

Hacking groups appear to largely still be in the reconnaissance phase of probing vulnerable systems to understand what kinds of data might be available to them, according to Jack Hughes, head of digital forensics and incident response for EMEA for Palo Alto Networks Inc.’s Unit 42 division.

He anticipates that there will be hacking victims because more than a third of the 22 million vulnerable systems that were exposed to the internet are running software that can’t be patched and protected, he said. Upgrading those systems will take months or longer, he added.

“This is going to have a very long tail,” Hughes said in an interview. “We expect this to be a problem for months going forward.”

One reason there’s no evidence of a wide-scale hack yet is because the flaws are hard to exploit in practice even with the help of AI, said cybersecurity researcher Marcus Hutchins in a LinkedIn post.

Uncovering New Flaws

The vulnerabilities at the heart of this issue had existed undetected for more than a decade before AI tools helped researchers from Alphabet Inc.-owned Wiz, OpenAI and the cybersecurity firm Calif discover them and work on a fix with the group that maintains that software.

This is known in the cybersecurity industry as responsible disclosure, and fixing the issues in an orderly way is generally straightforward and drama-free. This time, however, hackers appear to have obtained information about the vulnerabilities before the public disclosure and leveraged them in attempted attacks.

Amazon.com Inc. said that it had evidence of hackers attempting to exploit the security holes in the days before the public disclosure. Alongside the use of AI, the apparent leak of the information dramatically accelerated the timetable for security organizations to release the information and for PHP users to respond.

Oliver Simonnet, lead cybersecurity researcher at CultureAI, said that while AI is helping hackers get faster at identifying and exploiting security holes, defenders have the same tools and they are also making them faster at their jobs as well.

Clinton, who before Anthropic worked on the Google Chrome browser’s infrastructure security, predicts that AI models will advance to the point that they effectively end security vulnerabilities.

Software developers “everywhere will be using models for fractions of pennies on the dollar to stop shipping security bugs permanently,” he wrote on X. That will represent a moment that Clinton called “vulnzero,” meaning an era of perfectly secure software with zero security flaws.

Before that can happen, however, organizations will need to deal with a “vulnerability tsunami,” he wrote.

— By Jordan Robertson and Ryan Gallagher (Bloomberg News)