AI traffic is exposing the internet's rickety infrastructure, and buyers will pay for it

The constraint is that nobody holds the inventory

The announcement that went out this week told website owners to patch immediately or take their sites down. The vulnerability was not new. As Anthony Ha put it on TechCrunch's Equity podcast, it had existed in PHP for decades and nobody knew, and the trigger for disclosure was that researchers had found it and could exploit it, so the calculation became that others could too.

Kirsten Korosec's point is the one that should worry an IT director: a lot of people do not even realise their websites have PHP. She describes it as the core ingredient behind content management systems used throughout WordPress and Drupal, something like 70% of websites, tens of millions of them. An estate that believes it runs a managed CMS is still running PHP underneath it.

That is the constraint that decides how this lands. The patching decision is not difficult. Knowing what to patch, and who owns the box it sits on, is. Sean O'Kane's image for the bottom of the stack is that when you get down to the bone level of the internet it is held together with glue and dreams, and the glue is usually a component nobody put on the architecture diagram.

Two numbers, one direction, and no standard way to count

Fortune reports that Cloudflare, the internet security and performance company used by millions of websites, says the crossover happened in June, with bots generating 57.5% of webpage requests. Thales, the French data security group, dates the same crossover to 2023 and its Bad Bot Report puts bot traffic at 53% in 2026. Both agree bots outnumber humans. Neither agrees on when it started or how to measure it.

Rudy Yang, Pitchbook's enterprise and retail fintech analyst who wrote the firm's July report on agentic AI traffic, told Fortune there is no one standard way to measure bot traffic because no single provider sees activity across the entire web. His reading is that the missing pieces point the same way: there is more bot activity, and agentic AI activity is driving a lot of the browser activity being observed.

For a buyer, the disagreement is the finding rather than a reason to discount both. A vendor pitching bot mitigation will quote whichever figure flatters its product, and the honest position is that the direction is established and the magnitude is provider-dependent. Ask which measurement a supplier is using before it becomes a line in a renewal deck.

The PHP incident shows what an agent-speed disclosure looks like

The vulnerabilities were found by human researchers participating in OpenAI and Wiz's Hack the Planet initiative, though they used AI to exploit the bug, as TechCrunch corrected in its own text. Wiz is now owned by Google. The distinction matters because it is the difference between AI finding flaws and AI accelerating the exploitation of flaws humans found, and only the second is documented here.

O'Kane describes the disclosure rollout as contested, with people unhappy about it and what he calls a bit of chest beating in OpenAI's post. A couple of people got ahead of the embargo, which set off a race among the hacker set. Cloudflare chief executive Matthew Prince posted that he had already seen it escalate to millions of attempted hacks of PHP2Shell, and Korosec calls that post telling and pretty scary.

O'Kane's structural point is that normal vulnerability disclosure works because you can go to a company and tell it discreetly, and it stays in that bubble. When the affected software is as fundamental as this, there is no discreet route, because you cannot identify and reach every person who controls some part of a website. He notes Amazon also published a blog post on it, which is what a well-capitalised company with a team can do and what most estates cannot.

The consequence for an enterprise is that the window between public disclosure and mass scanning has compressed. Calif's write-up of PHP2Shell describes two PHP remote vulnerabilities, CVE-2026-109618 and CVE-2026-109619, addressed in the PHP October 2026 Security Release, and traces the underlying out-of-memory issue back to PHP bug #78876 reported in 2020 and an OSTIF and Quarkslab audit in 2024 where exploitation was not proven and the bug was not fixed.

Where the load actually breaks an enterprise

Korosec's second observation is link rot. Companies, governments and small neighbourhood sites are shutting pages down rather than patching, so material that was reachable stops being reachable. For an enterprise that is a supply-chain problem as much as a historical one, because the documentation, status page or regulatory notice an integration depends on can disappear without a change ticket.

The traffic side compounds it. If agents are generating the majority of requests, then capacity planning built on human session counts understates the load, and rate limits sized for people do not bind on software. Fortune's reporting does not put a cost on this and neither Cloudflare nor Thales publishes a figure for what bot traffic adds to a hosting bill, so treat any per-request cost claim as unpublished until a supplier shows its method.

O'Kane's framing of the risk is about multiplication. He asks how many times in the last decade one person maintaining a tiny repository as a hobby has caught something and saved the infrastructure from crumbling, then says to multiply that by a billion when people can task agents to go and do the finding themselves. The defensive side of that equation is not obviously staffed.

What it costs before and after

No source here publishes a price for absorbing agent traffic, and that absence is the finding. Cloudflare and Thales publish shares of requests, not bills. Fortune quotes analysts on measurement, not on spend. An IT director asking what this does to the budget should expect a quote-only answer from every vendor in the category and should ask what unit the quote is metered on.

The one number that is published and does move a bill is the share itself. If 57.5% of requests are bots on Cloudflare's measurement, then a plan priced per request is already majority non-human, and the question of whether agent traffic is billable, blockable or cacheable is a contract term rather than an engineering preference. The same applies to any integration that calls a third-party API metered by call volume.

The PHP side carries a different cost, which is the patch window. Taking a site down, as the disclosure advised for anyone who could not update immediately, is an availability decision with a revenue line attached, and the estates most exposed are the ones that did not know PHP was in the stack. That is an inventory cost before it is a licensing one.

The decision rule for the next renewal

Choose to renegotiate the contract if the provider cannot tell you what share of your requests are automated and how that share is billed. That is the single question that separates a supplier who has instrumented this from one who has read the same headlines. Cloudflare's 57.5% and Thales's 53% are different numbers because they measure different things, and a supplier that cannot say which thing it measures is guessing about your bill.

Choose to run the inventory before the next disclosure if you cannot answer today which of your public-facing properties run PHP and who patches them. Korosec's estimate that something like 70% of websites depend on it means the answer is probably yes somewhere, and the October 2026 release is the precedent for how much notice you get. The researchers found this one. The next one may be found by an agent tasked to look.

Choose to treat link rot as a dependency risk if any integration, runbook or compliance record points at a page you do not control. Korosec raised it as a consequence of sites closing rather than patching, and it is the cheapest of the three to mitigate. The expensive one is the bill that arrives because nobody asked what a request is worth when most of them are not people.

What is still unpublished

Neither Cloudflare nor Thales publishes a per-request cost for bot traffic in the material here, and Fortune does not put a figure on the commercial impact. Any number a vendor offers for what agent traffic adds to an enterprise bill should be treated as quote-only until it comes with a method. The published figures are shares of requests, and they are contested between two credible providers.

The PHP vulnerabilities have identifiers and a release date, which is more than most of this story has. CVE-2026-109618 and CVE-2026-109619 are documented, the October 2026 security release addresses them, and the underlying issue has a paper trail back to 2020. The bot-traffic numbers have none of that discipline yet, which is why the direction is safe to repeat in a meeting and the percentage is not.