PHP2Shell: AI-assisted attack compromises 25 servers and a government network

An attack that combines web and kernel vulnerabilities

Security researchers have published the details of an incident they describe as an AI-assisted attack, attributed to a threat actor allegedly linked to China. According to the analysis, the attacker exploited the vulnerabilities known as PHP2Shell to break into exposed systems and, once inside, chained a kernel privilege-escalation exploit to gain full control (root) over the compromised machines.

The episode, observed on October 25, 2026, is said to have affected 25 servers and a network affiliated with the U.S. Government. The report does not detail the exact entry vector or the specific PHP versions affected, but it does make the pattern clear: a web application vulnerability serves as the door, and a second flaw, this time in the operating system, allows escalation to the highest privilege level.

Why this matters to a small business

The first reaction of many small teams is to think that an attack of this kind only affects large organizations or state infrastructure. That is a miscalculation. Attackers do not choose their victims by size, but by exposure: a server with an outdated PHP version, an administration panel accessible from the internet, or an unpatched container is enough to start a chain like this one.

Moreover, the use of AI-assisted tools to automate the search for flaws and the construction of exploits reduces the time between the publication of a vulnerability and its exploitation. We have seen this recently in Altman's warning about old flaws that AI is uncovering, and also in the cases where Anthropic's AI has chained vulnerabilities in browsers and operating systems. The conclusion is the same: the reaction window is shrinking.

The technical lesson: chaining flaws is the norm

The PHP2Shell case illustrates a pattern worth internalizing. A successful attack rarely depends on a single vulnerability. The usual thing is a chain: a misconfigured or unupdated web application allows code execution; from there, a kernel flaw or poor permission assignment allows jumping from user to user and taking control of the server. That is why isolated measures are not enough.

At ForgeNEX we recommend reviewing the exposure surface with a chain mindset, not a loose-patch one. Some useful questions to start with:

• Do we have PHP applications exposed to the internet that do not need to be?
• Do the servers run the web code with an unprivileged user and with an updated kernel?
• Do we monitor the creation of unusual processes or privilege elevation?
• Do we know which PHP and operating system versions run in production and who maintains them?

What we recommend doing this week

There is no need to wait for an incident to act. These are the actions we prioritize in projects for small businesses and IT teams with limited resources:

1. Inventory and urgent patching. Locate all PHP installations, both on your own servers and in containers, and update to supported versions. Also review the kernel and service permissions.
2. Reduce exposure. Close administration panels, debugging tools, and services that do not need public access. If it is not used, it is not published.
3. Principle of least privilege. The web process should not run as root or with write permissions on sensitive paths. An application flaw should not turn into a total compromise of the host.
4. Detection, not just prevention. Log and alert on privilege elevations, unexpected command executions, and unusual outbound connections. Many chains are detected at the moment the attacker tries to escalate.
5. Segmentation. Separate web servers from databases and internal systems. If one server falls, the rest should not be within reach.

It is also worth reviewing CISA advisories on critical PHP flaws, which we have already discussed on this blog, and maintaining a tested response plan. AI speeds up attackers, but it can also help us detect and prioritize. The difference lies in having the fundamentals in order.

An unpatched server is not just a technical problem: it is a door that can become the entrance to the entire network.

For teams managing clusters or containers, it is also worth reviewing the hidden flaws that bring down the Kubernetes control plane, because the logic of escalation and persistence is similar. And if your small business is starting to use AI agents in production, it is worth reading the lessons about agents that escape before expanding their autonomy.

Source: The Hacker News. Analysis and adaptation: ForgeNEX.