The PHP security bug — and why your workforce is the control that scales

PHP2Shell was found by AI, leaked before its fix, and drew millions of exploit attempts within a day. The patch closes two CVEs. Only a workforce that can use AI closes the next hundred.

Facts current as of October 25, 2026

What is the PHP security bug?

Two things make this incident different from the hundreds of critical CVEs published every year. First, the flaws were found by AI. Researchers at Calif, working inside OpenAI’s Patch the Planet program, and independently at Wiz used frontier models to find bugs that had sat in PHP for over a decade. Second, the same AI capability that found them lowers the cost of exploiting them: within a day of disclosure, exploit attempts had reached the millions. That combination is the reason incidents like this will become more frequent, and the reason the enterprise response has to be a workforce response, not only a patching one.

What is happening with PHP2Shell right now?

PHP shipped emergency fixes a day early after details leaked, exploit attempts reached the millions within a day, and a further PHP advisory followed two days later. As of October 25, 2026, this is where things stand:

The two bugs are memory-safety flaws. CVE-2026-109618 is triggered when an out-of-memory error interrupts PHP while request data is only partially initialized. CVE-2026-109619 lets request state survive after its memory is freed, so a later request on the same worker can use it (Google Cloud).

The exposure is large. PHP runs roughly 70–80% of websites whose server-side language is known, including WordPress, Drupal, Joomla and Laravel. Researchers counted about 1.9 million internet-exposed hosts, and most servers that disclose a version run branches that no longer receive security updates (CyberScoop). For those, there is no patch to apply, only an upgrade or a shutdown.

Why will incidents like PHP2Shell happen more often?

Four forces are compounding:

1. AI is finding bugs that humans missed for years. The out-of-memory flaw behind CVE-2026-109618 was spotted in a 2024 audit but never fixed, because no exploit existed and maintainers could not agree on a fix (Wikipedia). AI models produced the exploit. OpenAI says PHP2Shell is one of several significant vulnerabilities it has already found, and that it expects similar findings to increase sharply in the coming months (CyberScoop). In the same month, researchers reported using a public OpenAI model to build a full Chrome exploit chain in six days (The Hacker News), an AI-found Linux kernel flaw that gives local users root (The Hacker News), and an AI-found flaw across HEIF image decoders (CyberScoop).
2. Exploits that used to require specialists are now cheap to adapt. Wiz’s researchers note that writing a working exploit for a memory bug across many PHP versions and configurations used to take rare expertise. Large language models can now adapt an exploit to a new version or memory layout and run the testing loop, which changes the economics of both discovery and exploitation (CyberScoop). Attackers get the same tools defenders do.
3. The window between disclosure and mass exploitation is now measured in hours. Fastly observed active probing across PHP applications within a day and put the gap between disclosure and repeatable exploitation at hours to a few days for a flaw of this severity (CyberScoop). Cloudflare logged tens of thousands of exploit attempts within 45 minutes of disclosure and millions by the next morning, and by October 23 botnet operators had joined in. A reliable exploit was still proving harder to build than expected, and confirmed compromises remained few: that gap is the window defenders get, and it closes with every iteration an attacker’s model runs (The Hacker News).
4. Vulnerability volume is already at record levels. Microsoft’s October 2026 Patch Tuesday fixed 1,413 vulnerabilities, more in one month than any prior annual total (CyberScoop). The PHP Foundation that had to ship the PHP2Shell fix has 13 full- and part-time employees (Wikipedia).

The net effect: more critical flaws, found faster, exploitable sooner, landing on security and IT teams whose headcount is flat. The bottleneck has moved from the vendor’s patch to the enterprise’s ability to find every affected system and act on it before the exploit is automated.

What turns a known vulnerability into a breach inside an enterprise?

PHP2Shell makes that visible: the fix existed on October 21, exploit attempts began the same hour, and the organizations still exposed days later are the ones that could not inventory, triage and update fast enough. Correlation One’s discovery work with enterprises in 2026 keeps surfacing the same four conditions behind that lag.

Each of these conditions existed before PHP2Shell. The incident is the first time many leaders will feel their cost in hours.

Why is AI enablement for the workforce one of the best security investments an enterprise can make?

A patch fixes one CVE. A WAF rule blocks one exploit pattern. A workforce that can use AI competently on inventory, triage, remediation and communication gets faster on all of them, and the gains compound as vulnerability volume grows. Three reasons the return is unusually high right now:

1. The licenses are already paid for. Most large enterprises have sanctioned AI tools deployed to tens of thousands of seats. One large financial services firm we spoke with had enabled Copilot Chat for more than 70,000 associates without a formal rollout. The period between tool availability and productive use is where the value is won or lost. Enablement converts a sunk cost into capability; it does not add a new line item of the same size.
2. The attacker’s advantage is a skills advantage. Wiz’s point is that AI reduced the specialist effort needed to turn a decade-old memory bug into a working exploit. The defender’s equivalent is an engineer who can use the same class of model to scan a codebase for affected versions, draft the upgrade across every image and pipeline, write the regression tests, and summarize exposure for the board, in hours rather than days. That engineer is trained, not hired; the market does not have enough of them to hire.
3. Governance only works at the keyboard. Incidents are when policy is most likely to be bypassed. An enablement program that pairs persona-based permissions with training, so that each tier of employee knows what they may do, with which tool, on which data, is the control that holds when a team is working at 2 a.m. to take a PHP 7.4 server offline. Written AI policy without trained behavior does not.

The programs that produced measurable results in our 2026 engagements shared a sequence: calibrate to real workflows, establish a measured baseline, build shared fluency, split into role tracks (engineering separate from corporate), redesign actual workflows in applied labs, then sustain with champions and governance built in. Security and IT teams belong in the engineering track, with incident response as one of the workflows redesigned.

What should enterprise leaders do about PHP2Shell this week?

1. Update every PHP installation to 8.5.12, 8.4.27, 8.3.36 or 8.2.35, then restart the PHP process. Include container images, CI/CD base images and vendor appliances, not only servers. Anything on 8.1 or earlier has no patch: upgrade it or take it offline, per CISA.
2. Set session.upload_progress.enabled=0 until the next routine PHP release. It is PHP’s listed mitigation for CVE-2026-109619 and its recommended action for the separate CVE-2026-115560 advisory of October 23.
3. Hunt back to October 17. Amazon’s telemetry shows attempts four days before disclosure. Check web-server logs for shells spawned by the PHP worker and unexpected files in the web root.
4. Inventory what you could not find quickly. Every PHP host that took more than a day to locate is a gap in asset visibility. Record it.
5. Time the response. Hours from advisory to full remediation is the metric that matters as disclosure-to-exploitation windows shrink. Most organizations have never measured it.
6. Ask which teams used AI during the response, and how. If the answer is “nobody” or “unofficially,” that is the enablement gap. If someone used an unsanctioned tool on production data under pressure, that is the governance gap.
7. Baseline AI fluency in security, IT and engineering before the next one. PHP2Shell will not be the last critical flaw found by a model this quarter. Build the engineering AI track, with incident response as a redesigned workflow, while the cost of the lag is still fresh.

Frequently asked questions

Correlation One delivers AI, data and cybersecurity workforce training for enterprises, governments and development programs. Since 2015 we have trained more than 500,000 professionals across 50+ countries and delivered programs for 500+ employer partners, measuring success in applied business outcomes, verified jobs and wages.