As the usage of artificial intelligence — benign and adversarial — increases at breakneck speed, more cases of potentially harmful responses are being uncovered.
Pixdeluxe | E+ | Getty Images
Government websites in Europe were taken offline on Wednesday as concerns spread about vulnerabilities in servers that host sites coded in PHP, the world’s most popular server programming language.
The U.S. National Security Agency warned millions of organizations to either implement patches immediately or take their servers offline. It’s the latest cyber scare in the artificial intelligence era, as vulnerabilities that previously took weeks or months for hackers to find can now be exploited in a matter of hours.
Matt Suiche, a researcher at cybersecurity startup Tolmo, says this level of urgency is “very new.”
“That’s not something that even existed a year ago,” he said. “The reason there is that sense of urgency now is really because everyone has AI-augmented capabilities, not only defenders but also attackers.”
What happened?
PHP, created in the 1990s, is used by just under 70% of all websites with known server-side programming languages.
In mid-August, researchers at AI firm Calif and Google’s Wiz identified vulnerabilities in all PHP installations serving web requests. For the discovery, Calif’s researchers used OpenAI’s GPT‑5.6 Sol, while the Wiz team used its Wiz Atlas security AI agent, Google’s Gemini 4 Argon, and GPT-5.6 Sol.
The vulnerabilities, dubbed “PHP2Shell,” allow for unauthenticated hackers to remotely execute code on any PHP server, letting bad actors potentially steal data and read sensitive information.
The PHP Foundation was intending to release patches for the bugs on Thursday. But news of the vulnerabilities leaked earlier in online forums, and then by Calif itself, which accidentally posted a proof-of-concept video on Tuesday that demonstrated an exploit chain.
Though the video was ultimately taken down, the patches were fast-tracked and released 24 hours early by the PHP Foundation, at 10 a.m. ET on Wednesday. Shortly before that, the NSA issued its warning.
By that point, code had already been posted to public repository GitHub, which is owned by Microsoft, claiming to be able to exploit PHP2Shell for any server.
Elizabeth Barron, executive director of The PHP Foundation, told CNBC on Wednesday afternoon that the organization is “diligently monitoring the situation,” but noted that there is currently no “evidence of any working exploits, only incomplete, misguided, and non-working generated ones.”
However, Suiche said his team at Tolmo was able to code a working exploit in less than a day.
“With AI, it was relatively quick to have a local reproduction,” he said, adding that before today’s powerful AI models it would have taken his team “multiple weeks” to produce it.
Amazon Threat Intelligence reported on Wednesday that it identified a PHP2Shell exploit attempt on Saturday on its MadPot telemetry network, a network dedicated to identifying digital decoy connection attempts.
‘Structural failure’
The vulnerabilities have been around for a long time. Two years ago, Mihail Kirov, then an engineer at Quarkslab and now a product security engineer at Tolmo, said he was part of a team that found a vulnerability with the same root cause as PHP2Shell.
Kirov said The PHP Foundation was alerted to the flaw in its systems at the time, but a fix was never rolled out. The foundation acknowledged in a blog post Wednesday that the delay was a “structural failure, caused by both the security team’s decision process and its lack of resources.”
That was before large language models were so pervasive. Calif’s researchers had new powerful technology at their fingertips. They found two vulnerabilities that were both rated a 9.2/10 in terms of security severity.
“They proved that today, in 2026, using large language models, things can go really, really fast,” Kirov said, about the Calif team. “One of the reasons for that is that LLMs allow you to explore so many different angles.”
Even as organizations apply patches to their vulnerabilities, they’ll likely never know if data was breached, experts say.
Neatsun Ziv, CEO of cybersecurity provider Ox Security, said that unless an organization had preinstalled an application security platform before a hack attempt, it wouldn’t be able to see suspicious requests made on its servers.
Suiche said a company would only know it was hacked if its valuable data was “being sold later on the dark web.” He said that’s the “reality of yesterday and is still the reality of today.”
PHP is a legacy server language, Suiche said, so many of its users may not have implemented sophisticated security measures.
He said that city governments and other agencies often use PHP and “haven’t updated their website in 10 years.”
“Those people have a lot of critical data,” he said.
Industry experts say the cybersecurity game has changed dramatically since Anthropic released its powerful Mythos model earlier this year. The Claude creator pre-released Mythos in April to a small number of companies and government agencies to ensure the cybersecurity measures were sufficient before wide-scale release.
Now, Mythos is available to many bad actors.
“The amount of time that it takes the bad guys, the threat actors, to actually find a new exploit and run it on the entire internet used to be weeks,” Ziv said. “It went down to days, and starting in the Mythos age, from April 2026, it went down to hours.”
#data #risk
Source link
