On 21 October 2026 the developers of PHP, the language that runs much of the web, released urgent updates for two critical vulnerabilities. They had been in the code for years.
They were found by two research groups, with the help of artificial intelligence. The same day Microsoft described an attack campaign carried out with AI agents, and LiteLLM disclosed that a version of its software had been distributed with malicious code. Several institutional websites were taken offline as a precaution.
Within a few hours, this news had found a name: Black October.
A label, certainly, but the episodes raise a question that perhaps can no longer be underestimated: if AI makes it possible to discover and exploit vulnerabilities ever faster, will we be able to fix them just as quickly?
What is Black October
Between 22 and 23 October the term Black October appears in coverage by Axios, The Guardian and The Washington Post, thanks in part to John Milinovich, who works on artificial intelligence at Figma and published a very interesting piece on 22 October.
For Milinovich, the various signs of runaway AI, the system breaches and the dozens of exceptionally severe cyber attacks concentrated within 48 hours would signal a sudden acceleration.
Now, let us be clear: very different dynamics sit under the same label. Confirmed attacks, vulnerabilities discovered before they caused damage, incidents disclosed in the same days but which began weeks earlier. To understand Black October, all these events must be separated.

PHP: a years-old flaw, found with AI
PHP generates the pages and handles server operations for a great many websites: according to W3Techs data reported by Reuters, around 70% of those whose server-side language can be identified. WordPress, Drupal, Joomla and MediaWiki use it.
The two vulnerabilities, CVE-2026-109618 and CVE-2026-109619, were given a severity score of 9.2 out of 10. Under certain conditions, they would have allowed an attacker to execute code on a server without credentials: reading confidential information, altering content and compromising how the site works.
According to OpenAI, the company Calif identified one of them using GPT 5.6 Sol as part of the Patch the Planet programme. Researchers at Wiz found and reported both, using several AI tools. Note, however: those flaws had been in the code for many years, and artificial intelligence showed that weaknesses considered hard to exploit could now become dangerous.
On 21 October the European cybersecurity agency ENISA, together with the European Commission, CERT-EU, Europol and other authorities, urged organisations to update their systems immediately. In many cases the shutdowns of institutional websites were precautionary. It is the first time in our history that a tool has corrected us humans.

Microsoft: automated attacks are already a reality
The same day, Microsoft Threat Intelligence published an investigation into Storm-0940, a group the company considers aligned with the interests of the Chinese state. According to Microsoft, the group used ARTEX, an artificial intelligence tool for security testing, to automate reconnaissance (access attempts and abuse of permissions in cloud systems), also passing through an application linked to the IT services provider ConnectWise.
The attackers reportedly tried to break into more than 2,250 customer cloud environments. In some operations, less than three hours passed between choosing the target and obtaining privileged access. Most attempts failed, but in some organisations the attackers maintained unauthorised access.
While the attribution to China remains an intelligence assessment by Microsoft, the central point here is the mechanism. An attack goes through several phases: choosing a target, gathering information, finding a way in, checking permissions, widening access. An artificial agent can chain them together and decide the next step on its own. For those defending a network, the time between the first suspicious signal and compromise is shrinking.
LiteLLM: when the risk comes from a trusted supplier
On 21 October LiteLLM, a platform used to integrate different AI models into applications, announced that version 1.107.0 had been published on its official channels with malicious code designed to steal credentials. It remained available only briefly.
There is no evidence to attribute the attack to an autonomous AI system. The case does, however, show the risk of the software supply chain , the chain of tools, libraries and suppliers used to build a digital product: a company can protect its own systems well and still import the problem by installing an update from a source it considers reliable.

Three scenarios for today
Artificial intelligence is a magnificent tool, capable of changing cybersecurity on three fronts. The machine analyses large volumes of code and finds combinations of errors that would take a person weeks. It helps to understand under what conditions a flaw can be exploited. And it repeats: an agent runs checks, interprets the results and carries on without every step requiring human intervention.
All this, however, opens the way to faster attacks. And so the time to react shrinks: a company can receive a fix and find itself facing automated attempts before it has installed it.
A wave of vulnerabilities is therefore plausible. Jason Clinton, head of security at Anthropic, has spoken of a possible vulnerability tsunami. Matthew Prince, chief executive of Cloudflare, has pointed to the risk that organisations will be unable to fix problems as quickly as they are found. The forecasts rest on real progress, but the outcome depends on maintainers’ resources and the speed of updates: the same wave could bring chaos, or the fixing of long-standing problems before anyone exploits them.
But the real risk, a somewhat more speculative one that is circulating widely online, is of course the fast takeoff. The scenario of Milinovich, Situational Awareness and AI 2027: growth in artificial capabilities so rapid that it leads to systems able to improve themselves. The true transition from artificial intelligence to superintelligence.
What to do now
The useful measures do not depend on which scenario comes to pass.
Knowing what the company runs on is now more essential than ever. An e-commerce site, a customer portal or a management system may depend on software whose functions are all the business knows about. You need to know which technologies are in use, including those managed by suppliers, and update them, prioritising systems exposed to the Internet. For PHP, the patched versions released on 21 October are 8.2.35, 8.3.36, 8.4.27 and 8.5.12. Updating WordPress or its plugins is not enough: the engine must be updated.
At that point, governance, credentials and permissions need strengthening. LiteLLM was compromised partly because of a credential that was never revoked; in the Microsoft case, cloud application permissions opened the way to other environments. What is needed is rapid revocation of exposed keys, phishing-resistant multi-factor authentication for sensitive accounts, and restorable backups. As we explain in our dedicated guide to IT governance, these are all decisions that involve several business functions, not just IT.
Finally, you need to decide in advance who decides. Who receives a report, who can isolate a system, who communicates during an incident.
The race between discovery and remediation
Black October exposed a problem across the entire digital infrastructure. Software written many years ago still underpins economic activity, public services and platforms we use every day, and today it can be examined with far more powerful tools.
Finding a flaw, however, is only the beginning. The affected installations must be identified, the update prepared, its compatibility checked and the fix rolled out. This work runs through companies, developers, suppliers and public administrations, each with different resources and timescales.
How quickly will we be able to fix, compared with how quickly artificial intelligence will let us discover? A significant part of the security of digital services depends on that gap. And perhaps of our modern history.
Domande frequenti
What is Black October 2026?
An informal expression that circulated in October 2026 to describe a series of vulnerabilities, attacks and security incidents, some of them linked to artificial intelligence. It does not refer to a single attack, nor to an officially classified crisis.
Did artificial intelligence cause the Black October attacks?
In some episodes the use of AI is documented, as in the campaign described by Microsoft. In others, AI helped to discover the vulnerabilities, as in the PHP case. There is no evidence that all the incidents can be attributed to artificial systems.
Which sites are vulnerable to the PHP flaws?
Those running the PHP versions listed in the security advisories; the actual risk also depends on the server configuration. Site owners should check the software they run and install the updates issued by the PHP project.
Can artificial intelligence carry out cyber attacks on its own?
Agents exist that can automate several stages of an attack: Microsoft has documented their use in real campaigns, and Anthropic has published experimental results on models capable of developing exploits. Performance depends on the tools available, the protections in place and operating conditions.
Does Black October mean we are close to AGI?
No. The documented events concern programming, analysis and cybersecurity capabilities, and are not enough to establish the imminent arrival of artificial general intelligence.
Fonti e riferimenti
- John Milinovich, Black October
- Johana Bhuiyan, Black October: cybersecurity experts sound the alarm over new era of AI-enabled hacking
- Herb Scribner, The internet’s hidden security nightmare and how AI exposed it
- PHP Development Team, Security advisory: Critical remote code execution vulnerabilities, update now
- OpenAI, Patch the Planet: risposta globale alle vulnerabilità critiche scoperte dagli agenti
- Sagi Tzadik, Shir Tamari, Nir Ohfeld, Tomer Schwartz, PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet
- ENISA, Joint statement on critical PHP vulnerabilities
- Microsoft Security Research e Microsoft Threat Intelligence, AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise
- Oliver Jensen e Mateo Wang, The attack chain behind the v1.107.0 LiteLLM release
- Andrew Fasano, Marius Fleischer, Cole McFaul, Robert Xiao, Tripp Gallagher, GLM-5.3 and the spread of advanced cyber capabilities
- Leopold Aschenbrenner, Situational Awareness: The Decade Ahead
- Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean, AI 2027

