Talk Around Town: 'Black October' May Be the Week AI Broke the Internet — and Got It Broken Right Back

What the Hollerin' Is About

Well, butter my biscuit and call me alarmed — the week of October 20–24, 2026 handed the internet security world two gut-punches at the same time, and folks are still spitting dirt trying to figure out how bad it got. First came what Wiz Research calls the PHP2Shell disclosure: a pair of critical memory-safety holes in PHP that, according to Wiz Research, could let any random uninvited guest run whatever code they please on a vulnerable server without so much as a howdy-do login. Then, like a second possum falling out of the same tree, Socket researchers published findings that the release pipeline for LiteLLM — a widely used Python library for connecting to AI models — had been quietly hijacked and laced with malware before anyone noticed.

Security professionals have taken to calling the whole ugly cluster 'Black October,' according to Axios, though that same outlet noted that many organizations appear to have patched or mitigated fast enough to dodge catastrophic damage. That's a relief, sure, but it don't make the week any less terrifying to look at in the rearview mirror.

What Is Actually Known: The PHP2Shell Situation

Here's what multiple independent sources — CNN, TechCrunch, Axios, Qualys, Wikipedia's dedicated article, Wiz Research, and Recorded Future's Insikt Group — all agree on: two vulnerabilities tracked as CVE-2026-109618 and CVE-2026-109619 exist in PHP versions 5.4 through 8.5, they are rated a flat-out terrifying 9.2 out of 10 under CVSS v4.0 by PHP's own developers, and security firm watchTowr pushed the rating all the way to a perfect 10.0. Wiz Research says it found the flaws and reported them through coordinated disclosure before the October 21 public release. Wikipedia's PHP2Shell article notes the vulnerabilities were independently found by Ryan Austin of California-based firm Calif — reportedly working through an OpenAI program called 'Patch the Planet' — and by Wiz Research as a separate discoverer.

Because PHP powers something in the neighborhood of 70% of websites worldwide — a figure cited by Wiz Research and echoed across CNN, TechCrunch, and Axios — the blast radius concept here is, as we say down home, wider than a barn door. Platforms running on PHP include WordPress, Drupal, Joomla, and Wikipedia's own MediaWiki engine, according to multiple corroborating sources. CNN reported that government websites in multiple countries temporarily went dark following the disclosure. Both the US National Security Agency and Europe's cybersecurity agency, as reported by CNN and Axios, urged organizations to patch immediately or take affected systems offline.

Now here's where your narrator has to pump the brakes and be honest: Recorded Future's Insikt Group stated plainly that confirmed real-world compromises from actual exploitation had not been established as of October 21, distinguishing between observed exploitation attempts and verified damage. Other outlets framed the situation with considerably more alarm. So the frog is definitely on the hot stove — we just ain't confirmed it's cooked yet.

What Is Actually Known: The LiteLLM Supply Chain Attack

Socket, the firm that discovered the LiteLLM compromise, published a detailed technical account stating that attackers slipped malicious code into LiteLLM version 1.107.0 on October 21, 2026. According to Socket, that poisoned release was designed to harvest cloud and AI credentials, and the injected files were arranged so that Python could execute the payload during normal interpreter startup — meaning, per Socket's analysis, an application didn't even have to directly call LiteLLM for the nastiness to run. That's about as sneaky as hiding a catfish in somebody's Sunday hat.

CyberPress and Socket both report that the attackers had actually breached LiteLLM's GitHub organization eleven days earlier, on October 10. According to Socket, the initial foothold came from a GitHub App master key that had been accidentally committed to a public repository on May 27 and never revoked — sitting out there like a house key taped to the front door. Socket says the attackers eventually obtained owner-level OAuth tokens, giving them enough control to tamper with the release pipeline itself. Microsoft Security Research and The Hacker News published independent technical analyses corroborating the supply chain compromise. The exact number of downstream users affected during the roughly two-hour window the malicious version was live has not been publicly quantified, though sources variously cite approximately 23 million weekly downloads or figures as high as 92 million monthly downloads depending on measurement window.

What Remains Unverified and Disputed

A few things are still murkier than a pond after a tractor pull. On PHP2Shell's severity, there is a genuine disagreement: PHP's own developers assigned a 9.2 CVSS v4.0 score, while watchTowr rated it a full 10.0 — not a rounding difference, a meaningful gap in how dangerous these organizations judge the flaw to be. On the discovery timeline, Wikipedia and some sources place the initial find in August 2026, while at least one outlet references July 2026 as when Calif's researcher first used OpenAI tooling; OpenAI itself has not publicly confirmed which specific model was involved. On LiteLLM's download figures, sources are quoting numbers ranging from about 23 million weekly to 92 million monthly, a spread wide enough to matter when you're trying to gauge exposure.

Most critically for the headline-vs.-reality gap: Recorded Future's Insikt Group drew a clear line between reports of exploitation attempts targeting PHP2Shell and any confirmed evidence of successful, damaging real-world compromises as of late October 21. The actual damage tally for both incidents is still an open question, and anyone hollering that the internet definitely got torched this week is getting a little ahead of their headlights.

Analysis: AI as Both Hound Dog and Fox in the Same Hen House

This is analysis, not settled reporting — so take it accordingly. What makes 'Black October,' as Axios dubbed it, genuinely remarkable is the double-edged story it tells about AI's role in security. On one side, Wiz Research says AI-assisted research was central to finding PHP2Shell, a pair of bugs that had apparently been hiding in one of the most widely deployed programming languages on the planet for years. That's the hound-dog side: AI sniffing out buried problems that human researchers might have kept circling for another decade.

On the other side, the LiteLLM attack illustrates that AI developer tooling has become a high-value soft target for supply chain attackers, according to Socket's analysis. LiteLLM is infrastructure that sits between developers and large language model APIs, meaning a compromised version could theoretically harvest credentials from a broad swath of AI-adjacent software projects. The attackers reportedly got in not through some genius exploit but through a forgotten key left sitting in a public repo — which is the security equivalent of leaving the barn door open because you got distracted by a shiny thing. The convergence of these two events in a single week, whether coincidental or not, is what has the security community talking, and it's the kind of talking worth listening to even before all the facts are nailed down.