For much of the last month, many of the cybersecurity experts I speak with and follow have been posting as if they were very scared. Many have warned of an AI-powered “black swan” event, have been alarmed by the number of vulnerabilities suddenly discovered, and by the apparent power of AI agents and, more broadly, large language models to find and exploit previously unknown bugs.
The rhetoric ranges from worried but measured (things are about to get weird) to mundane advice like updating your software and making sure you have two-factor authentication, to vague publication on AI-based attacks, on things like “the Internet was nice during it really lasted Prepper level terror.
To give you an idea of what I see, Matthew Green, a professor at Johns Hopkins and one of the most respected cryptographers in the world, job “We’re at the part of the movie where you have to throw everything you can in the path of the zombies and hope it slows them down,” and “so many people are discovering that the the whole internet is dry tinder and the only thing keeping it from burning is the fact that the teenagers haven’t found the matches yet,” Green is alluding to the fact that hackers and researchers who drive AI tools and agents are getting better at finding and exploiting all kinds of bugs in important systems. I’ve also seen many AI employees, as mentioned, are preparing thoroughlysuggesting people withdraw money from their bank accounts, disconnect from the internet, buy shelf-stable foods, stockpile water, put money in gold and cash, and buy emergency hand-crank radios.
And all these messages came Before People using AI-based tools have discovered vulnerabilities in PHP, the programming language that powers much of the Internet. In recent days, major vulnerabilities or issues have been discovered in eMudhra (an Indian entity that provides website security certificates), Node.js (an alternative to PHP), Gitlab (a software development platform), Plex (media server software), Rail (a cloud provider), Open Code (a coding agent platform) and several other companies and services, although, understandably, none of them have received as much attention as the vulnerabilities in PHP, called PHP2Shell, that researchers by Wiz, CaliforniaAnd OpenAI discovery thanks to AI.
Matthew Prince, CEO of Cloudflare, a service that has a broad overview of the general health of the Internet, job“In the 10 hours before the public announcement of (PHP2Shell), on the Cloudflare network we saw thousands of exploit attempts. This increased to tens of thousands in the 45 minutes after the announcement, to hundreds of thousands in 2 hours, and to millions now.”
So much is happening, and so quickly, that it’s been difficult for me — for me who has been reporting and editing stories about security and hacking for a decade — to fully understand what’s going on, to get a sense of what might happen next, or to predict where it’s going to go. And I think that’s largely where we are right now, as I write this.
But it has become clear that AI, broadly defined, has fundamentally changed both offensive hacking and defensive cybersecurity. This is not because AI has become sentient, nor because humans have lost control of it, nor because “swarms” of “malicious agents,” but because AI is proving something we have always known: that much of the software that runs the Internet, our devices, our accounts, and our lives is vulnerable, or outdated, or has security holes that humans never noticed, or was installed by someone at a company that left. company six years ago and have been replaced four times and no one knows how to find them now or how to find them. update it. This does not mean that all vulnerabilities discovered through AI were obvious or insignificant; qualified hackers and researchers, piloting AI tools, seems to be better than humans alone to find vulnerabilities. But, crucially, AI and large language models have made the brute force of the Internet extremely easy; If every software is a door, AI makes it extremely easy to turn the knobs on each door to see if it’s unlocked, even on buildings that humans have never really bothered to try to open. What we’re learning at an alarming rate is that not only are some locks easy to pick, but in some cases the door is wide open or there’s a giant hole in the wall. The recent wave of people using AI to hack their Internet of Things devices to run new software shows that, in some cases, it takes virtually no expertise to use AI to hack something.
During this time, we have seen the rise of “abliterate” AI models designed not to say no and open weighting models that can be used for penetration testing and potentially offensive hacks without the guardrails that exist in many commercially available business models.
In a blog post explaining the PHP vulnerability discovery, OpenAI said the results highlight “how AI can uncover security vulnerabilities with Internet-wide impact, but the same capabilities can also help attackers quickly develop offensive tools and exploits.” OpenAI essentially argues that it did all of this research ethically and that its the discovery opened a “defender’s window” this allowed governments and critical infrastructure to prepare for the bug’s disclosure. “Researcher finds a bug, ethically discloses it to the company, the company fixes it” is the preferred course of action in cybersecurity, but AI shortens the window for all that, gets more people to do more research more easily, and speeds up both finding and fixing bugs. Whether these patches are actually being deployed more quickly by individual companies is another much more complex question. In many cases, the answer is probably no.
AI also makes it all a lot more unpredictable. The security company Chainguard has been criticized for breaking the PHP vulnerability disclosure embargo, as it released patches for the vulnerabilities hours before their official disclosure. The critics, including the PHP Foundationsaid this could have allowed attackers to figure out how to exploit the issues before website owners had a chance to protect their systems. In a post on LinkedIn, Dan Lorenc, CEO of Chainguard, said that his the company independently found the same bugs by typing “find bad vulnerabilities in PHP” into a frontier AI model. Amazon, for its part, said it I saw attempts at exploitation before official disclosure. These facts raise the question of whether the bug was exceptionally easy to find, whether the AI models that initially discovered the vulnerabilities incorporated that information into the model, making it easier for someone else to find, or whether there was a leak along the way from someone connected to the disclosure process. Any option is cause for concern.
After the PHP2Shell vulnerabilities were discovered, some computer security experts started calling it “Black October” and highlighted the fact that the National Security AgencyTHE Prime Minister of Japanand various governments around the world asked people to immediately update their websites and services, as well as remove services that could not be fixed. 4chan, several Wikimedia servicesAnd Archives.org have all been temporarily taken offline so they can be updated.
As The Atlantic already suggestedOne of the biggest impacts of PHP2Shell might be the fact that many legacy websites cannot or will not be updated, may be taken offline proactively, could possibly be hacked, and may never come back. The link rot could be horrible and a lot of the story could be lost.
For such a high-profile incident, most of the real impacts so far have been related to proactively applying patches or removing potentially insecure websites, and we have not seen immediate, widespread exploitation of the bug. Various low-effort exploits developed by LLM have been abandoned online; many of them won’t work. A blog post titled “There is no easy exploit for this PHP vulnerability,” written by Elias Taylor, which quickly went viral and was dissected on Hacker Newsnotably by people pointing out that the site eliastaylor.com had been registered recently and that it was not clear who Elias Taylor was. 404 Media reached Taylor through Signal; he said “I’m real and on a late train.”
“A lot of people are fooled by AI into thinking they have an achievement when they don’t, because they can’t read or understand what the LLM gave them,” Taylor told 404 Media. “This doesn’t seem to be as exploitable as the other ‘stop the world’ vulnerabilities, i.e. log4j, heartbleed. But patch your servers, it’s still a vulnerability. Please don’t paraphrase me by saying ‘there’s no vulnerability’ for any reason. There is, and people should patch.”
Dan Guido, CEO of Trail of Bits, a highly respected cybersecurity company partnering with OpenAI on the Planet Patch initiative, posted that his company had created “a working exploit” and that “Will Tan built it in 1 day with Codex, drawing on about 10 years of research into system security and vulnerability. AI makes exploit development faster. Knowing how to turn a bug into a shell always requires expertise. AI makes that expertise more productive.”
It should also be noted that Taylor another blog post is titled “Everything Connected to the Internet is Going to Be Hacked”, which includes the lines “software engineers have spent the last 40 years writing vulnerable, crappy software. LLMs are easily above the bar needed to break into said software.”
We don’t yet know exactly what the final impact of all this will be, or where we will be in an hour, a week, a year or 10 years. But AI is changing the way cybersecurity and hacking work. This makes finding and exploiting bugs faster and easier. We’re about to find out just how insecure the Internet is.
Gn bussni
