The week of October 20, 2026, stress-tested every assumption the software industry holds about vulnerability disclosure, and most of them failed. Two critical remote code execution bugs in PHP — CVE-2026-109618 and CVE-2026-109619, collectively branded PHP2Shell — were discovered independently by California security firm Calif (under OpenAI's Patch the Planet program using GPT-5.6 Sol) and Israeli-U.S. firm Wiz (using Gemini 4 Argon, now part of Google Cloud). The bugs carry CVSS scores of 9.2 and affect an estimated 70% of the internet, from Wikimedia Foundation to DeviantArt to UK government sites and Japanese banking systems. The disclosure timeline was a cascade of failures compressed into 17 hours. A Hacker News user named petecooper posted about the CVEs, apparently unaware an embargo existed. The PHP Group rushed its announcement forward from Thursday the 22nd to Wednesday the 21st, but the patch still wasn't ready. Calif leaked a video of the exploit in action — complete with dancing anime girl and background music — while security firm Chainguard published fixes six hours before the embargo lifted. By the time the official patch shipped, the vulnerabilities were already being exploited in the wild. Governments in the UK, EU, Japan, and the US issued urgent advisories, some overblown, sending sysadmins into panic. The structural shift here is not the bugs themselves — PHP has always been a target-rich environment. It's the time compression. One of the two CVEs (109618) was actually identified during a PHP Foundation security audit in 2024, funded by Germany's Sovereign Tech Agency, but was deprioritized because nobody could build a working exploit from the theoretical description. LLMs closed that gap instantly. Frontier models can now take a vague description of a theoretical vulnerability, trace diffs in a codebase to locate the relevant patch, reverse-engineer the underlying flaw, and produce a functional exploit — all in a workflow that takes minutes, not the weeks or months that human researchers historically required. This means the entire coordinated-disclosure model — where researchers find bugs, notify vendors, and give them time to patch before going public — is operating on borrowed time. The moment any signal of a security patch enters the public record, AI agents can race backward from the fix to the flaw faster than most organizations can deploy updates. As the article's author puts it: the announcement of a security patch should now be treated as the last possible date the full exploit is not being distributed. The flood of non-working exploit slop produced by less capable agents doesn't change this calculus — one working exploit among ten thousand duds is all it takes. The PHP2Shell crisis didn't happen in isolation. On the same day, Mozilla blocked eight intermediate certificates from Indian certificate authority eMudhra after a pseudonymous researcher discovered their private keys were laughably weak — numbers like "115", "56", and "4" instead of hundred-digit values. eMudhra has since announced it will exit the SSL/TLS certificate business entirely. Cloud provider Railway disclosed that hackers accessed and decrypted environment variables from their platform, potentially targeting AI inference providers. GitLab disclosed two separate CVSS 10.0 RCE vulnerabilities (CVE-2026-97229 and CVE-2026-110264) related to regex parsing, following yet another 10.0 from the previous month (CVE-2026-85706). The pattern across all four incidents is not coincidence — it's convergence. The attack surface of internet infrastructure is enormous, the tools to probe it are now automated and cheap, and the defensive ecosystem still operates on disclosure timelines designed for a world where exploit development required deep human expertise and significant time investment. That world ended sometime in the last twelve months. The eMudhra case shows that even without AI, basic competence failures persist in critical trust infrastructure. The Railway breach suggests adversaries are specifically targeting AI inference supply chains, adding a recursive dimension: AI tools finding vulnerabilities in systems that serve AI tools. The question is no longer whether AI changes the security landscape — it demonstrably has. The question is whether the institutions, processes, and incentive structures that govern software security can adapt to a world where the exploit window has collapsed from weeks to hours. The PHP Foundation ignored a known vulnerability for two years because no human could weaponize it. An LLM did it in minutes. That gap between institutional speed and adversarial capability is now the central risk in software security, and nothing about the current response suggests anyone has a plan to close it.
