PHP, a certificate authority, and an AI proxy all failed the same week, same reason.

Three unrelated systems broke the same week, for the same underlying reason, and nobody planned it that way.

On October 21, 2026, PHP shipped emergency patches for two critical remote-code-execution flaws (CVE-2026-109618, CVE-2026-109619), rated CVSS 9.2, in code some reporting says had been there for fourteen years. Greg Brockman, OpenAI's president, credited OpenAI's own models with finding it and called it one of the most severe bugs in open source software in recent memory — though other reporting says PHP's own maintainers had already flagged part of the flaw back in 2024, and multiple independent teams were circling the same bugs by the time it went public. Government agencies across the US, UK, EU, Canada, Australia and Japan told organizations to patch immediately; the FBI and NSA told anyone who couldn't to take the affected systems offline instead. Some did — government and news sites in several countries went down as a precaution.

The next day, eMudhra — a certificate authority most people have never heard of — announced it is exiting the global SSL/TLS certificate business entirely. A researcher had recovered the private keys for eight of its intermediate CA certificates on October 17; eMudhra's own statement blamed “a flaw in the code” that generated a weaker second key pair during the certificate ceremony. Chrome and Firefox distrusted the root behind roughly 30,800 certificates. The SSL/TLS line was only about 1% of eMudhra's global revenue. The company killed it anyway, rather than try to re-earn trust in it.

The same week, LiteLLM, a widely used open-source AI model proxy, shipped a compromised release (1.107.0) on PyPI carrying a credential stealer and backdoor. The entry point, per the maintainers' own account, was almost too on-the-nose: a GitHub CLI OAuth token had gotten embedded in an AI coding tool's responses, which LiteLLM's own request and response logs then captured. An attacker found those logs, harvested the token, and used it to compromise the release pipeline. It was the second LiteLLM PyPI supply-chain compromise in roughly seven months.

None of these are the same bug. They're the same failure mode: something was sitting somewhere, reachable, holding a secret or running code that nobody could take back once it was out.

Anthropic, whose own model (Claude Opus 4.6) has found and validated more than 500 real high-severity open-source vulnerabilities, has started saying the quiet part out loud. Its coordinated-disclosure policy now paces submissions to what maintainers can actually absorb, rather than sending large volumes of findings at once, and its research team has written that “industry-standard 90-day [disclosure] windows may not hold up” against bugs an AI can find faster than a human can fix them. Finding the bug stopped being the hard part. Reaching the machine running it, and getting someone to act before anyone else does, is the actual bottleneck now — and that doesn't change just because the finder is a model instead of a person.

FOKL's own tools are built to run sandboxed, client-side, on the one device they're ever installed on. No server holding a secret, no proxy logging a token, no certificate chain to distrust. Not because the engineering is smarter than the PHP core team's or eMudhra's key-ceremony engineers' — they weren't careless, the surface they're defending is just enormous. It's because there's nothing deployed anywhere for an attacker, or a model, to eventually go find.

This was one week. It won't be the quietest one of the year.