Originally published on Substack: https://paragonaimag.substack.com/p/ai-can-now-hack-the-internet-and — republished here for developers and builders, with the same 7-step checklist.
Watch the full video: https://www.youtube.com/watch?v=hUg3-eb-A3A — Listen to the podcast episode: https://redcircle.com/shows/cbce814a-d451-4913-a015-fe05b4d0e059/episodes/d97fb13d-72d3-4c04-9645-9e094996921e
By Paragon AI MAG — 30 October 2026. Every fact below is backed by sources and was checked several times before release, so it reaches you with the fewest errors and flaws. We never claim to be entirely error-free.
The week governments switched their own websites off
In August 2026, AI models auditing the internet surfaced a severe vulnerability that had sat undetected in the web's plumbing since 2012. On 21 October 2026, an emergency patch went out worldwide. Within hours, governments across Europe began deliberately pulling their own national portals, identity services and municipal sites offline. They chose to shut down entirely rather than risk leaving those portals exposed to an automated attack.
We are no longer waiting for AI-driven cyber threats to materialise in some distant future. The machine-speed arms race is happening right now.
One flaw, millions of websites: the shared brick
To understand how a single flaw caused global panic, look at the building materials of the web. A programming language called PHP is the mortar holding together platforms like WordPress and Wikipedia, assembling web pages on the server before they ever reach your browser. PHP powers 69.3% of identifiable server-side websites (W3Techs, 22 October 2026). That does not mean 70% were vulnerable — but it does mean this shared brick is everywhere. One cracked brick design puts every house built with it at risk.
The bugs, nicknamed PHP2Shell, are two vulnerabilities (CVE-2026-109618 and CVE-2026-109619) with a 9.2 severity rating out of 10. An attacker could execute commands without a password, gaining direct access to files, databases and encryption keys.
Fortunately, the defence got there first. Independent researchers, acting as auditors, deployed AI models — OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon — to comb through the codebase, finding this decade-old hole before it could be weaponised at scale. The California firm Calif found a critical flaw on 1 August 2026; Wiz found both vulnerabilities independently weeks later and reported them to PHP on 24 August. The quiet two-month window that followed is how coordinated disclosure works: researchers tell the vendor privately, a fix is built in quiet, and only then is the flaw announced together with the patch.
Even so, the exposure prompted the US National Security Agency to issue an urgent directive: patch immediately, or disconnect. The Netherlands, Finland and France took that literally, with preventive-maintenance shutdowns across major state networks — the Dutch national cyber centre, the Overheid.nl portal, the DigiD identity service, the Port of Rotterdam, Finland's public web services, France's Interior Ministry sites. Turning off your own digital infrastructure is a massive, disruptive step. It was a desperate, necessary move to win a race against attackers that defenders didn't even know they were running.
How an AI finds a hole
Historically, finding a flaw like PHP2Shell required a human to manually read lines of code, looking for tiny logical mistakes. Today, an AI model does the same job like a proofreader who never sleeps, instantly referencing every bug pattern ever published. The models also use fuzzing: throwing millions of odd or unexpected inputs at a program, at machine speed, to see exactly what breaks it. Instead of random guessing, the AI intelligently selects which strange inputs are most likely to force an error.
This is a dual-use capability. It is the exact same tool used by a locksmith testing the deadbolt on your front door, and by a burglar methodically checking every house on your street in a single night.
The sheer volume of flaws this technology unearths is visible in Google's Chrome 156 release of 20 October 2026, which patched 353 distinct security flaws in one release, six of them rated Critical, several surfaced with the help of AI assistants — including one found by an Anthropic researcher with Claude's help and one in WebRTC reported by OpenAI Codex Security. Two honest notes: 353 patches is a massive number, but it is not a record (Chrome 150 had 382), and Google did not state these critical flaws were being actively exploited in the wild. The story is the pace of discovery, not a single scary number.
The core asymmetry of cyber warfare remains: the defence is required to secure absolutely everything, while the offence only needs to find one single unpatched server to break in.
55 years of dumb worms — and the worm that learns
For the last 55 years, malicious software mostly relied on brute force. Worms like Creeper (1971), Morris (1988) and Mirai (2016) were essentially unthinking programs executing a single repetitive trick until a human engineer intervened. Even devastating outbreaks like WannaCry (2017) spread through a rigid, fixed process.
In October 2026, Anthropic's Frontier Red Team published a study showing how these automated programs are adapting to operate independently. Researchers William Zhou and Daniel Freeman purchased 15 off-brand Android TV boxes, the kind you can buy for about $55, and networked them together to run a distributed AI model. In a controlled firewalled network, they deployed a patient-zero AI worm. It successfully hopped from one device to the next through an open debugging port, setting up its own inference server on each newly infected host. Then came the critical shift: the AI copies began passing notes to the next generation about their failed strategies. By learning from the previous copy's mistakes, their successful setup rate jumped from 45% to between 61% and 84%.
To be clear, no fully autonomous AI worm has escaped into the real world yet, and the break-in steps in this lab test were kept intentionally trivial — an unauthenticated port, no password. The result is an upper bound on real-world success, not a live outbreak. But the shift in tactics is already visible outside the lab: attackers recently breached the developer platform Railway to steal AI API keys, treating these keys exactly like stolen credit card numbers. The objectives of cyber warfare are shifting — attackers break into platforms and steal API keys specifically to hijack hardware and fuel their own massive AI compute operations. An API key is the new password, except it never sleeps and nobody types it.
The 7-step checklist: secure your digital life today
You cannot control a global machine-speed arms race, but you have absolute control over closing the doors these automated systems are actively looking for.
- Turn on automatic updates everywhere — your phone and browser, but critically your router and smart TVs. Those devices are computers now, and they hold enough processing power to be targeted.
- If you own a website, update now — WordPress core, plugins, and your PHP version to patched builds (8.2 through 8.5) this week. Back up first. If your host cannot patch, move hosts: PHP versions older than 8.1 have no patch for this flaw at all.
- Use a password manager to guarantee a strictly unique password for every single site you use. One breached account should never unlock the rest of your life.
- Turn on two-factor authentication or passkeys across your accounts. Prioritise your email first, because your inbox acts as the master key that can reset everything else.
- Treat unexpected security alerts as hostile until you verify them by manually logging into the official app or website — never through a link in the message. Establish a family code word to verify any panicked call about a financial emergency.
- For developers and builders: rotate your API keys after any breach news, set hard spending caps on your accounts, and never paste your keys into random online tools.
- When a service announces a breach, change your password on that specific site first, then on any older accounts where you might have reused it, and closely monitor your linked payment methods.
Executing these seven steps transforms a soft digital footprint into a hardened target. Automated AI threats are designed for efficiency. They will simply bypass you for easier prey.
The honest close
For now, the defence is keeping pace. The massive PHP flaw was discovered by AI, reported safely, and patched before it could trigger a widespread disaster. What to watch next: whether AI-found flaws keep arriving at October's pace, whether any self-replicating AI system is ever seen outside a lab, and whether the next big flaw is reported by a defender first — like PHP2Shell — or found by an attacker first. That single difference decides whether the story ends in a patch or in a breach.
Paragon AI MAG tracks revolutionary AI research and explains what it means — before it changes everything.
Sources
Reuters / AFP coverage of the October 2026 PHP disclosures; Wiz and Calif research blogs; OpenAI's disclosure note; Google Chrome Releases (Chrome 156, 20 October 2026); Anthropic Frontier Red Team, "AI models may be capable of self-replicating without access to a datacenter" (12 October 2026); W3Techs server-side language statistics (22 October 2026); CrowdStrike reporting on AI-powered attacks on South Korean financial institutions; Railway incident reporting.
Subscribe free to Paragon AI MAG so you never miss the daily video and podcast: https://paragonaimag.substack.com/
Top comments (0)