[AINews] OpenAI’s Patch the Planet program is 1 of 2 parties to discover PHP2Shell vulnerabilities in core PHP affecting 70% of the web, using GPT-5.6 Sol; coordinated response over 2 months

What it is

PHP2Shell is the name researchers gave to two PHP-engine memory-corruption flaws, CVE-2026-109618 and CVE-2026-109619. The page is an AINews digest (20-21 October 2026) that reconstructs the event from social posts. It is secondary reporting, not an advisory. It contains no PHP Foundation release notice, no patch commit and no fixed-version matrix.

How it works

The researchers' description places the defects in PHP's internal handling of out-of-memory conditions and fatal errors, not in a CMS plugin or application authentication logic. That is why waiting for plugin developers cannot fix it. Researchers report two outcomes: pre-authentication RCE demonstrated against WordPress, Drupal and Joomla, and "Heartbleed-style" remote memory disclosure against any PHP application. The digest warns that these are researcher-reported results and do not show that every deployment has an identical RCE path.

Preconditions / Affected

The page gives no affected or fixed versions, no CVSS vectors and no deployment prerequisites. It lists these as open gaps, along with what leaked before release and successful exploitation versus scanning. A viral claim of about 1.5 million unfixed servers (attributed to Censys) and reports of shutdowns at Wikipedia, Archive.org and Dutch institutions are uncorroborated. The page does not say whether those outages were precautions, patching, unrelated failures or compromises.

What to do

The page states the following:

• Cloudflare deployed WAF rules, free tier included, covering known exploit vectors. Its CTO called them a stopgap and told operators to patch at 14:00 UTC (23:00 JST on 21 October 2026). Cloudflare reported "early attacks" but gave no counts or compromise details.
• NSA guidance is to prepare for an immediate update, and to take systems offline if timely mitigation is not possible.
• OpenAI published a repository with a Codex skill that walks operators through checking whether a site is affected. The digest says this is an aid, not a replacement for an advisory or package verification.
• The digest's suggested incident sequence: 1. Get the authoritative PHP and distribution advisories. 2. Inventory exposed runtimes, including managed hosting, containers, appliances and background workers. 3. Apply temporary protections where justified. 4. Deploy authenticated fixed packages. 5. Verify the running build and restart processes. 6. Investigate exploitation separately, since updating does not show that credentials or data were never exposed.
• Check that WAF coverage actually applies. Alternate ingress paths and origin exposure must be verified, not assumed.

Caveats

• Discovery was independent. Calif researcher Ryan Austin found the flaws under OpenAI's Patch the Planet using GPT-5.6 Sol, and Wiz reportedly found them with Gemini 4 Argon. OpenAI says "to our knowledge" defenders found them first, which does not rule out earlier attacker knowledge.
• The two-month coordination timeline comes from participants (OpenAI, Brockman, Wallace), not an audit. The patch reportedly required a major engine change.
• Reports that details leaked early come from The Rundown. The page does not establish what leaked or whether the schedule changed.
• The "15-minute to weaponization" figure was a prediction, not a measurement.
• Two-factor authentication does not stop an unauthenticated engine bug.
• Japanese breach statistics from SemiAnalysis are separate incidents, not tied to PHP2Shell.
• The digest itself says the severity comparison to log4shell and Heartbleed is debated.
• The fetched text was truncated at 60,000 of 61,854 characters. The cut falls in unrelated Reddit recap material, so the PHP2Shell section appears complete.