“Black October”: AI finds PHP flaws, attackers strike within minutes

What's going on

October 2026 has seen a string of security incidents that experts have lumped together under the label “Black October.” The biggest stir came on October 21, when emergency updates were released for two critical vulnerabilities in PHP, a programming language that runs on millions of websites. The flaws were found with the help of AI. What makes them unusual: according to Wiz researchers, exploiting them required an extremely complex chain of steps, at the edge of what human researchers can realistically pull off.

One important caveat up front: according to Fox News, “Black October” describes a series of security emergencies, not a single coordinated hacking campaign. exchange4media also stresses that there is no confirmed evidence of one coordinated attack on the entire internet.

The PHP flaws: “PHP2Shell”

The two vulnerabilities are tracked as CVE-2026-109618 and CVE-2026-109619 and are collectively called “PHP2Shell.” Under certain conditions, attackers could use them to run malicious code on vulnerable servers without logging in first (remote code execution, meaning code runs remotely on someone else's machine). Successful attacks could expose sensitive data or hand criminals control of systems.

According to Reuters (cited by exchange4media), PHP powers just under 70% of websites where the server-side language can be identified. Some experts estimated, per the Guardian, that up to 70% of all sites were at risk. Fox News clarifies that this doesn't mean 70% were hacked or that every PHP site is vulnerable. Affected are PHP installations that process web requests, and exploitation may require complex technical conditions.

In a blog post, Wiz called the flaws “among the most severe vulnerabilities discovered in PHP, and arguably in software history.” The underlying bugs reportedly date back decades. According to exchange4media, OpenAI and Wiz, Google's security subsidiary, were involved in the AI-assisted research.

The attack wave and the response

According to Cloudflare CEO Matthew Prince, there were signs of thousands of exploitation attempts even in the 10 hours before the disclosure. In the 45 minutes after, there were tens of thousands. Within two hours, hundreds of thousands. Later, millions.

Agencies and companies reacted drastically. Government sites in the Netherlands, France, Denmark and Finland were taken down as a precaution, as were government and private sites in the US and UK. Japan also saw temporary disruptions to bank and government services. The US NSA advised all PHP operators to apply the patches immediately. Such access could help attackers steal sensitive data, disrupt services or dig deeper into networks.

Top executives sounded the alarm:

“I no longer believe we can patch all the bugs.” – Sam Altman (per the Guardian)

Jason Clinton, Anthropic's deputy chief security officer, warned of a “vulnerability tsunami” for which there isn't enough time to patch critical infrastructure. Salesforce CEO Marc Benioff called it “Code Red,” and Microsoft CEO Satya Nadella said, according to exchange4media, that he hadn't seen anything comparable in a long time. Prince expects an “incredibly tricky” two years, with tools that search the web for vulnerabilities with infinite patience.

More incidents involving AI tools

The PHP flaws are only part of the picture:

• LiteLLM: The open-source platform that connects applications to multiple language models reported on October 21 that its release pipeline had been compromised. Version 1.107.0 contained malware that steals credentials and runs commands remotely. Affected versions were removed. Users should check whether credentials were exposed.
• OpenCode: On October 17, the AI coding platform reported unauthorized access to a legacy Console system, which may have leaked stored API keys for third-party AI providers. Around 1.4% of Console users with connected providers were affected. Founder Colin Parsons reported about 50 billing reports: $281 at OpenAI and $168 at Anthropic. OpenCode linked the incident to a large-scale distillation attack on AI models (a model is retrained on another model's outputs), but hasn't publicly explained exactly how the usage came about.
• Microsoft and Storm-0940: On October 21, Microsoft described how the China-linked group Storm-0940 used the open-source framework ARTEX, an AI-assisted penetration-testing tool. The targets were hundreds of organizations. On several occasions, attackers gained privileged cloud access in under three hours. From October 10 to 14, Microsoft counted more than 151,000 failed logins against 327 accounts in eight organizations, up to 2,900 per minute (password spraying, where one password is tried across many accounts). There were no successful logins in this campaign. The group has also been linked to an incident at ConnectWise.
• Others: Critical flaws in Node.js and GitLab, plus an incident at India's eMudhra involving credentials for an intermediate certificate authority, which could enable forged certificates. India's CERT-In warned on October 15 about AI agents exceeding their permissions and noted that the time from breach to damage has shrunk from days to hours.

What you can do now

According to exchange4media, the PHP flaws have been patched, but systems remain vulnerable until the updates are applied. If you run your own PHP servers, update right away. If you use AI services via API, review your keys regularly, set spending limits and revoke keys immediately if you suspect a problem. If you used LiteLLM 1.107.0, treat your credentials as potentially compromised.

Perspective

The case shows both sides of AI in security: it finds bugs humans missed for decades, and it speeds up attacks at the same time. What's new is above all the speed, since exploitation attempts were already under way before disclosure, according to Cloudflare. A critical note: much of the data comes from vendors and executives who also have business interests, and the “70% at risk” claims are rough estimates. It remains unclear how many attacks actually succeeded and exactly how OpenCode keys leaked. Keep in mind that the incidents haven't been shown to be coordinated, despite the dramatic name.