Login
Inside the cyber flaw scramble that brought the AI threat to life
The exposure was immense, with AI multiplying the risks — and yet time-tested global cybersecurity procedures appeared to be functioning. Then suddenly, it all broke down.
The decision by a cybersecurity company on Wednesday to prematurely release information about a flaw in the popular PHP computer programming language left much of the web exposed to hackers, triggering a panic that brought to life months of warnings of how the AI-powered era could cause widespread damage.
The flaws in PHP, discovered using AI tools through a months-long process coordinated by OpenAI, caused cyber agencies across Europe, the U.S., the U.K., Australia and Canada to post urgent warnings Wednesday. Numerous governments took websites offline as a precaution.
While no major hacking damage was immediately apparent, the way the incident unfolded over 72 hours showed how easily cybersecurity processes can break down and the outsized consequences of cyber risks in the AI era.
The man who set off the chain reaction with the early release — Chainguard’s CEO Dan Lorenc — quickly apologized in a LinkedIn post. But on Thursday, in an interview with POLITICO, Lorenc said he’d shared the information ahead of time because the flaw was already being exploited. Though others disputed that, Lorenc said he believed he’d seen enough evidence to invalidate an embargo set up to allow time to find a fix for the flaws.
While he disputed that his actions would have caused problems for other defenders, Lorenc conceded that the evidential threshold in these coordinated cyber defense systems is a “grey area.”
The revelation of the PHP flaw brings out into the open what cyber officials have been bracing for since the arrival of the most powerful tools of artificial intelligence, which not only can be used to discover vulnerabilities but also to exploit them at scale.
The cybersecurity world’s long-established protocols helped prevent disaster scenarios, in which hackers would have successfully exploited the vulnerabilities to take control of numerous websites and access their data. But now the scale and end-of-days potential of AI-driven cyber risks are visible for the world to see.
This “is what happens when AI turns up a flaw that’s been hiding for years,” said Eylam Milner, co-founder and chief technology officer at cybersecurity start-up Echo. The vulnerabilities, he added, are “about as serious as it gets for the web.”
Origin story
Understanding how this week’s emergency unfolded offers a glimpse at how companies and authorities are working to stave off attacks — and how much risk they face going forward.
The origins of the incident can be traced back as far as April when top frontier AI firm Anthropic unveiled Mythos, leading some in the cybersecurity world to worry that systems would quickly be overrun by advanced artificial intelligence, with potential for catastrophic events to ensue.
But while researchers did find more flaws in computer systems in the months after Anthropic and other AI firms like OpenAI released those models, they had little impact in the real world. Even the Hugging Face incident, a widely discussed hack that saw OpenAI’s autonomous AI agents go rogue during an internal security evaluation, related more to the level of control AI companies exert over their models than their power to break into systems worldwide.
That changed this week after researchers unveiled the vulnerability in PHP, a coding language estimated to be integrated into some 70% of the web.
The process by which information about flaws is released among cybersecurity officials and affiliated companies is crucial. Through a longstanding practice known as coordinated disclosure, researchers and companies that identify flaws can find a fix and share it confidentially with agencies and others before releasing to the public, meaning they don’t hand malicious hackers the gift of a hole without a fix.
In the case of the PHP flaw, that process worked well, but not perfectly. Information about the bugs started to circulate ahead of the planned release time. Then, according to the PHP Foundation, Chainguard “unilaterally broke the embargo,” which “disregarded weeks of work by many parties toward a coordinated disclosure.” PHP Foundation executive director Elizabeth Barron told POLITICO by email that while it was “disappointed” with Chainguard, it hopes to “rebuild trust with them in the future to ensure we can continue to collaborate effectively.”
After the leak— which meant information about the bugs was in the open before the fix was ready — confusion quickly ensued, with government bodies and private companies choosing to temporarily shut down their websites rather than risk potentially damaging breaches to their data and the sites themselves.
Early Thursday, Matthew Prince, the CEO of Cloudflare, an internet services company that is widely used by websites to block cyberattacks, giving it a high level of visibility over what hackers are doing, said in a post on X that it had seen “millions” of attempts to exploit the flaw.
Cloudflare saw thousands in the 10 hours before the flaw was publicly announced, then tens of thousands in the 45 minutes after the announcement, hundreds of thousands within two hours, and now millions, Prince said.
Amazon’s cloud division also said the flaw is “now being broadly exploited in the wild.”
One senior U.S. government cyber official said that federal authorities have been overwhelmed — and not just by PHP. The same day the world was racing to address PHP, they were also dealing with bugs in node.js, a widely used JavaScript framework, Fortinet, GitHub and Oracle.
“The fire department is fighting multiple fires at once,” the official said, adding: “The good news is they’re all one-story structures on fire, which the fire department deals with all day, every day.”
More to come
Dauntingly, OpenAI said there’s more on the way. “These bugs are among several significant vulnerabilities we’ve already found and are working to fix, and we expect similar findings to increase sharply in the coming months,” the company said Wednesday in a statement.
Matthijs van Amelsfort, head of the Dutch cybersecurity agency, said in an interview that the job of keeping computer systems safe has been transformed by the advancement of AI. “As soon as the frontier models came out, that’s exactly what we said, this will be the trend,” he said. “And actually that’s happening now.”
Thomas Caspers, vice president of Germany’s cyber agency BSI, concurred: “We expect that [to be] happening more and more during the coming months.”
AI also ramps up the risks once the flaws are out in the open. Cyber agencies’ recommendation that organizations take their websites offline was “really a drastic measure,” Caspers said. “We wouldn’t have done that a few years ago or so.” But because AI has made it quicker and easier to exploit flaws, “we had to do that,” he said.
At the same time, the use of advanced AI to detect the flaws may create a new level of protection against malicious hackers looking to take advantage of them. Caspers said that though his German agency has also seen attacks, it has not seen any successful use of the flaw by hackers.
“It seems a little bit more complicated than expected to exploit,” he said.
Despite the fact that AI offers a new way to discover flaws, the work involved in keeping systems protected is more quotidian, with cyber agencies urging organizations to install the fix shared by the PHP Foundation, in a process known as patching.
“There are three things that you have to do right now,” said van Amelsfort from the Dutch agency. “And that’s patch, patch, and patch.”
John Sakellariadis contributed to this report.
