deadrisk

Analysis

Finding Flaws Got Cheap. Fixing Them Did Not.

Last week's worst PHP flaw was first spotted in 2024 and sat unfixed for two years. Outside reports this summer, some made with AI, forced the fix. The work fell to a foundation with one full-time employee and 14 contractors.

NETTENBRBRKCVLTMSFT
Coverage intelligence, not investment advice — methodology.

Two years, then 71 days

The worst of last week’s PHP flaws was found by a person in 2024 and left unfixed until outside reports, some made with AI, forced the issue. The PHP Foundation’s account gives the history. In 2024 PHP had its first external security audit in years, carried out by Quarkslab and paid for by Germany’s Sovereign Tech Agency. While fixing one of Quarkslab’s findings, a core developer contracted by the foundation found the first signs of the memory flaw behind CVE-2026-109618. An exploit was theorised but none was shown, and other issues came first. The security team drafted several competing fixes, each with costs, and the work stalled and restarted over months. The foundation’s own verdict: “the delay was a structural failure, caused by both the security team’s decision process and its lack of resources.”

This summer the outside reports arrived. Security firm Calif began auditing PHP on 30 July under the Patch the Planet programme. On 11 August a Calif researcher reported code execution without a login, from the same root cause. OpenAI reported that root cause independently; Wiz and OpenAI added attack paths on 24 August and 18 September. The WSJ reported that OpenAI’s GPT-5.6 Sol helped find the flaw in August. Earlier in the year the foundation had also scanned PHP with Anthropic’s Claude Mythos. From 11 August, it says, the fix was its security team’s top priority, with several people on it full time.

The fixes shipped on 21 October, 71 days after Calif’s report. They were due on 22 October, but the CVE numbers began circulating online on 20 October and the release moved up a day. Wired reported that AWS saw at least one actor try the flaws on 17 October, four days before the patch. Cloudflare’s chief executive said attempts on its network rose from thousands to millions after disclosure. Reuters reported European government, school and hospital sites taken offline while their owners caught up.

Supported fixes cover PHP 8.2 and newer. The security team also posted minimal source patches for 7.4, 8.0 and 8.1; anything older depends on vendors. On 23 October php.net added an advisory for CVE-2026-115560, a code-execution path that survives the patch in an unusual session setup, and suggested turning off session upload progress until the next release.

12
TEXXR edges for the PHP Foundation, all logged 21 to 25 October 2026
None before 21 October. PHP itself has 20 edges since 2007: 12 from 2007 to 2021, and 8 in the four days from 21 October

More reports, the same few people

Two of the foundation’s twelve edges are about its staff. Reuters reported one full-time employee and 14 contractors, now fixing a flaw the WSJ said puts up to 70% of websites at risk. The PHP security team has published 15 CVEs since 1 July and worked through more than 100 reports that were invalid, duplicates or could not be reproduced. “AI helps the team validate reports and group reports sharing the same root cause, but a person makes every call and works on every patch,” the foundation wrote. It expects “the number of reports will keep growing”, and says engineers who know the whole codebase are “too few”.

Other maintainers had said much the same. In April Bloomberg reported that Mythos added to concerns about maintainers’ workloads. In May curl’s Daniel Stenberg reviewed five Mythos findings and judged three false positives and one “just a bug”. A week later Linus Torvalds said duplicate reports had made the Linux security list “almost entirely unmanageable”. On 1 October Google stopped taking product-flaw reports to its open-source bounty programme over a flood of invalid AI submissions.

The largest vendors face the same arithmetic. Microsoft fixed a record nearly 200 flaws in June, then about 972 in September, taking its 2026 total by September to 2,760, more than double the year before. Within a week it shipped an emergency fix for problems those patches caused.

The attack side is cheap

The same week showed what the tools can cost an attacker. NorForge’s Snorre Fagerland told Reuters that a Chinese-speaking operator’s AI agents tried more than 1,200 sites and broke into websites of more than 100 companies, Deere and SharkNinja among them, for about $5,000 of tokens. Microsoft said China-linked hackers used an AI testing tool, ARTEX, to compromise ConnectWise in a campaign aimed at hundreds of organisations. On the defending side, Anthropic said Claude flagged 493 issues across 25 open-source projects, and gave no overall count of how many have been fixed.

Where this meets the forecast

Our build-out forecast has an AI-warfare world. In it, AI takes a growing share of about $2.7 trillion of military spending, is served first in a chip shortage and pays half again per token. Cyber is where the record already shows military interest: The Intercept described a nascent Army Cyber Command programme, ECHOTRIBBLE, for AI agents that would move through networks on their own. Last week’s numbers point to a different cost structure, though. A hundred breaches for $5,000 is a small token bill. The scarce input in the PHP case was not compute but the few engineers able to choose a fix. If cyber conflict runs on that pattern, its chip bill stays small and its repair bill does not. The forecast counts the first and has no line for the second.

What the vendors can sell

That repair work is where our cyber coverage looks for revenue. Cloudflare shielded sites before the patch, but it gave the rule to free customers too, so the response shows reach, not sales. Tenable sells exposure ranking: which systems to patch first when patches arrive faster than staff can apply them. Rubrik and Commvault sell recovery for the breaches that get through. Each still faces the test set last week: paid use, renewals and cash. Cheap findings create work. They do not yet show who is paid to do it.

Sources

Dates are the publisher’s where the page or URL shows one, otherwise TEXXR’s.

  • The PHP Foundation, account of the security release (“Today’s PHP security release and how we’re strengthening PHP security”), 2026-10-21. TEXXR: https://texxr.com/1297522
  • PHP, news archive: the 21 October security advisory, and the 23 October informational advisory on session upload progress (CVE-2026-115560). TEXXR: https://texxr.com/1297518
  • Cloudflare, WAF protection for the PHP vulnerabilities, 2026-10-21. TEXXR: https://texxr.com/1297519
  • AI just found a bug that puts up to 70% of the world’s websites at risk — Wall Street Journal, 2026-10-22. Original: https://www.wsj.com/tech/cybersecurity/ai-just-found-a-bug-that-puts-up-to-70-of-the-worlds-websites-at-risk-fd0902ea · TEXXR: https://texxr.com/1297520
  • AWS analysts say at least one actor attempted to exploit the PHP flaws on October 17 — Wired, 2026-10-22. Original: https://www.wired.com/story/hackers-exploited-the-php-vulnerabilities-days-before-they-were-patched/ · TEXXR: https://texxr.com/1297523
  • Websites in Europe taken offline over the PHP flaws — Reuters, 2026-10-21. Original: https://www.reuters.com/technology/websites-taken-offline-over-php-risks-2026-10-21/ · TEXXR: https://texxr.com/1297521
  • Matthew Prince on PHP2Shell exploit attempts before and after disclosure — X, 2026-10-23. Original: https://x.com/eastdakota/status/2113079352682570051 · TEXXR: https://texxr.com/1297623
  • PHP Foundation patches are limited to PHP 8.2 and newer — Wired, 2026-10-23. Original: https://www.wired.com/story/this-is-not-a-drill-2-php-vulnerabilities-are-setting-the-internet-on-fire/ · TEXXR: https://texxr.com/1297625
  • The PHP Foundation operates with one full-time staff member and 14 contractors — Reuters, 2026-10-23. Original: https://www.reuters.com/technology/php-foundation-under-funded-with-one-full-time-staff-member-director-says-2026-10-23/ · TEXXR: https://texxr.com/1297733
  • Mythos adds to concerns about workloads for open-source maintainers — Bloomberg, 2026-04-17. Original: https://www.bloomberg.com/news/articles/2026-04-17/anthropic-s-mythos-adds-strain-on-cybersecurity-teams-facing-ai-threats · TEXXR: https://texxr.com/1167286
  • Daniel Stenberg on Mythos’s five curl findings — daniel.haxx.se, 2026-05-11. Original: https://daniel.haxx.se/blog/2026/05/11/mythos-finds-a-curl-vulnerability/ · TEXXR: https://texxr.com/1168745
  • Linus Torvalds says duplicate AI bug reports made the security list “almost entirely unmanageable” — The Register, 2026-05-18. Original: https://www.theregister.com/security/2026/05/18/linus-torvalds-says-ai-powered-bug-hunters-have-made-linux-security-mailing-list-almost-entirely-unmanageable/5241633 · TEXXR: https://texxr.com/1169099
  • Google freezes product-flaw submissions to its OSS bounty programme over invalid AI reports — Tom’s Hardware, 2026-10-04. Original: https://www.tomshardware.com/tech-industry/artificial-intelligence/google-suspends-part-of-the-oss-vrp-bug-bounty-program-due-to-an-influx-of-invalid-ai-submissions-product-vulnerability-submissions-ended-october-1 · TEXXR: https://texxr.com/1296272
  • Microsoft’s June 2026 Patch Tuesday fixes a record nearly 200 flaws — Krebs on Security, 2026-06-10. Original: https://krebsonsecurity.com/2026/06/a-record-breaking-patch-tuesday-for-june-2026/ · TEXXR: https://texxr.com/1170583
  • Microsoft’s September 2026 Patch Tuesday fixes a record ~972 vulnerabilities; 2,760 in 2026 so far — Ars Technica, 2026-09-09. Original: https://arstechnica.com/security/2026/09/microsoft-patches-a-record-972-vulnerabilities-112-of-them-critical/ · TEXXR: https://texxr.com/1286684
  • Microsoft ships an emergency fix for problems caused by September’s patches — The Verge, 2026-09-15. Original: https://www.theverge.com/news/995302/microsoft-out-of-band-windows-11-update-fix-issues · TEXXR: https://texxr.com/1287119
  • How one hacker used AI agents and a few thousand dollars to break into scores of sites — Reuters, 2026-10-19. Original: https://www.reuters.com/legal/government/how-one-hacker-used-ai-agents-few-thousand-dollars-break-into-scores-sites-2026-10-19/ · TEXXR: https://texxr.com/1297331
  • Microsoft says China-linked hackers used AI tool ARTEX to compromise ConnectWise — Reuters, 2026-10-21. Original: https://www.reuters.com/technology/chinese-linked-hackers-compromise-connectwise-target-hundreds-global-attack-2026-10-21/ · TEXXR: https://texxr.com/1297649
  • Claude flagged 493 security issues across 25 open-source projects — Anthropic, 2026-10-20. Original: https://www.anthropic.com/research/attack-surface-reduction · TEXXR: https://texxr.com/1297556
  • Leaked documents detail ECHOTRIBBLE, a nascent Army Cyber Command agentic AI programme — The Intercept, 2026-10-19. Original: https://theintercept.com/2026/10/19/army-autonomous-ai-agents-arcyber/ · TEXXR: https://texxr.com/1297362
  • TEXXR entity edges for PHP (20) and the PHP Foundation (12), queried 2026-10-26. These are extracted relationships, not report counts; the foundation’s staffing edges repeat one Reuters report across two days.
  • DeadRisk’s build-out forecast: the AI-warfare inputs (military spending, priority in a shortage, price premium).

Disclosure: this piece was drafted with Claude, a model made by Anthropic, whose Mythos and Claude models appear in it.