orange tomato
Page 1/10

WHEN SYSTEMS FAIL

A companion notebook to the deep dive — runtime flaws, breach paths, and the boundaries where the fix actually lives
RUNTIME
→
WHO HOLDS IT
→
WHO ACTS
  • The working question: when systems fail, who has to act?
  • Two layers, one shape — a flaw in the PHP runtime, and a breach that travelled through an AI tool's integrations.
  • In both, the fix lived one layer below where the reader looks.
  • Stopgaps buy time. Only the holder closes the flaw.

Deep dive: When Systems Fail — figures dated at inspection, bounds travel with every number.

orange tomato
Page 2/10

THE STACK

SITE · PLUGIN · THEME
→
PHP RUNTIME
→
SERVER · HOST · IMAGE
  • PHP2Shell: two unauthenticated RCE flaws in PHP's own memory handling (CVE-2026-109618, CVE-2026-109619), rated 9.2.
  • Both fire before application code runs — nothing your code does can stop them.
  • Updating WordPress or Laravel does not update the PHP beneath it.
  • The fix arrives through the server package, the host, or the container image.
  • WordPress shipped no update — there was no WordPress-layer fix to ship.
Ask: who controls your PHP binary — host, image maintainer, distribution, or you?
orange tomato
Page 3/10

THE TIMELINE

17 OCT: FIRST PROBES
→
21 OCT: PATCH + DISCLOSURE
→
MILLIONS / HOURS
  • Amazon spotted attempts 4 days pre-patch; Cloudflare shields up ~9h before announcement.
  • Leak forced a one-day-early release (21 Oct, not 22nd).
  • Post-disclosure arc (Cloudflare-sourced): thousands → tens of thousands in 45 min → hundreds of thousands in 2h → millions.
  • Governments went dark as precaution: NL, FI, FR, UK 30+ services, JP, Wikipedia, Archive.
  • Most back within days; FR / UK / FI restorations never published.

Counts are Cloudflare's own network — a measure of pace, not a census.

orange tomato
Page 4/10

STOPGAP vs FIX

WAF RULE — BUYS TIME
≠
BINARY UPDATE — CLOSES IT
  • Cloudflare: request + response signals in one transaction (first managed use).
  • AWS: rule group v2.3 then v2.4 (default), 29 Oct EOL guidance added.
  • Both vendors, same sentence: patching is the only complete fix.
  • A running worker serves old code until recycled — restart every worker.
  • Distro packages backport: version string lies, changelog tells the truth.
A firewall rule filters the traffic while the binary stays vulnerable.
orange tomato
Page 5/10

THE EOL GAP

8.2 – 8.5: PATCHED
→
7.4 / 8.0 / 8.1: SOURCE BACKPORTS
→
5.4 – 8.1: NO UPSTREAM FIX
  • Fixed: 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35. PHP 8.1 EOL Dec 2025.
  • PHP team published minimal source backports for 7.4, 8.0, 8.1 only.
  • Ubuntu ESM to PHP 5.5; Debian current branches + paid legacy; RHEL streams uneven.
  • Community backports for dead versions carry no standing.
  • Managed hosting: you cannot fix it — confirm somebody else did.

Managed bargain, stated honestly: no control, no action needed — if you know your side of it.

orange tomato
Page 6/10

THE BREACH PATH

CONTEXT.AI
→
WORKSPACE ACCOUNT
→
VERCEL ENV
→
SECRETS READ
  • April 2026: attackers entered Vercel through a compromised third-party AI tool.
  • Tool → employee Google Workspace → Vercel account → environment → env vars.
  • Taken: variables not marked sensitive. Untouched: published npm packages.
  • No model was hacked. No prompt injection. The tool was a doorway, not a culprit.
  • Closed 16 Oct; outside responders, law enforcement, vendor engaged.
Treat every unmarked variable as exposed. Rotate before deleting anything.
orange tomato
Page 7/10

THE NUMBERS, WITH LIMITS

  • 21% of 602 breach orgs reported an AI model/app incident (IBM 2026; Ponemon-conducted, IBM-sponsored; full method gated).
  • Denominator: orgs already in a breach study — not all companies.
  • Two 27%s: cloud misconfig + connected app/API compromise. Reported conditions — never added, never causal.
  • 92% lacked IBM-defined "proper" access controls. "Proper" undefined; proves no single control's power.
  • AI-specific causes listed too: poisoning, prompt injection, inversion.

Read as reported, not counted. The limits are the finding's frame, not footnotes.

orange tomato
Page 8/10

THE PATTERN

SITE OWNER → HOST
·
ADOPTER → INTEGRATOR
  • Each failure lived one layer below the reader's gaze; each fix names its holder.
  • Each stopgap bought time without closing: WAF vs binary; hygiene vs trust chain.
  • Each pace set the terms: millions of attempts in hours; secrets read before notice.
  • OWASP excessive agency: too much function, too much permission, too much autonomy.
  • NIST API protection (upd. Mar 2026): check authorization where the action happens.
A prompt telling the model to behave enforces nothing by itself.
orange tomato
Page 9/10

WHAT YOU CAN INSPECT

  • Who controls your PHP binary — and is the patched build installed, workers restarted?
  • Are your branches still supported? Whose backport covers you if not?
  • WAF on? Expiry noted — a window, not a wall.
  • Each AI tool: which identity, reaching which files / APIs / actions?
  • Does the connected system check permission, or trust the model?
  • Can you see what the integration touched — and revoke it?
  • Secrets marked sensitive? Unmarked ones rotated?

No headline number required. Each line names a holder — several times, you.

orange tomato
Page 10/10

SOURCES & METHOD

  • PHP advisory + release records; PHP Foundation 21 Oct post; Cloudflare + AWS posts with dated banners; CISA catalog entries.
  • Vercel incident bulletin (closed 16 Oct 2026); IBM public study materials; OWASP LLM06:2025; NIST SP 800-228 upd1.
  • BleepingComputer, Kaspersky, national outlets for the shutdown timeline.
  • Open items, dated: 115560 fix; FR / UK / FI restorations; mass-breach evidence; IBM method.
  • No Orange Tomato article cited as a source. Evidence review: distinct pass, not independent.

All figures carry bounds: Cloudflare-sourced counts, reported-not-counted shares, vendor-claimed verifications, single-worker proof.