[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 claims

PHP2Shell is two memory-corruption bugs in the PHP engine, CVE-2026-109618 and CVE-2026-109619. The researchers place the root cause in PHP's internal handling of out-of-memory conditions and fatal errors, not in any plugin or application auth logic. The demonstrations include pre-authentication remote code execution against WordPress, Drupal and Joomla. The disclosing researcher (@sagitz_) separately claims "Heartbleed-style" remote memory disclosure against any PHP application.

The newsletter cautions that these are researcher-reported results. They do not show that every deployment has the same exploitable RCE path. Because the fault is in PHP core, waiting for plugin or CMS updates does not fix it.

Evidence

OpenAI and participants (Eric Wallace, Greg Brockman, @itszn13) describe about two months of private remediation. The work included a major PHP-engine change, hardening with the PHP Foundation, and advance notice to governments. These are participant accounts, not an audited timeline. OpenAI says no attacker knew of the bugs "to our knowledge," which is not proof.

The patch shipped at 14:00 UTC on 21 October 2026 (23:00 JST). Before release, Cloudflare deployed WAF rules covering known exploit vectors for all customers, including the free tier. Its CTO called the rules a stopgap that were updated as "early attacks" appeared. The post gives no counts of requests, victims or successful compromises.

The NSA advised preparing for an immediate update and taking systems offline if timely mitigation was impossible. Several further claims are uncorroborated in the source:

• The Rundown reports that details leaked before the patch.
• A viral compilation cites about 1.5 million unpatched servers, attributed to Censys.
• The same compilation reports shutdowns at Wikipedia, Archive.org, Asana, GOG and European government sites.

Google says Gemini 4 Argon found more than twice as many vulnerabilities as baseline models in Wiz testing. It discloses no dataset, budget or false-positive accounting.

How to use it

The page gives no fixed-version matrix, CVSS vector, distribution package IDs or patch commit. Operators need the PHP Foundation and distribution advisories to know what to install. OpenAI published a repository containing a Codex skill that walks operators through checking exposure. The newsletter calls it an assistance workflow, not a replacement for vendor advisories.

The recommended incident sequence has six steps:

1. Obtain the authoritative PHP and distribution advisories.
2. Inventory every exposed runtime, including containers, appliances, background workers and old apps.
3. Apply temporary WAF protection where justified, and check alternate ingress paths and direct origin exposure.
4. Deploy authenticated fixed packages.
5. Restart affected processes and verify the running build.
6. Investigate past exploitation separately, because patching does not show that credentials or data stayed safe.

Enabling 2FA does nothing against this unauthenticated engine bug.

The issue also carries other items with adoption details:

• Anthropic API version 2026-10-21 removes temperature, top_p and top_k from Messages. New major SDKs default to this version, and 2023-06-01 remains supported. Teams upgrading SDKs should review this change.
• Recursive Flint is an Apache-2.0 coding MoE with 39.8B total and about 2.5B active parameters and 262K context. It needs roughly 24GB locally and is free on OpenRouter at launch. Its reported scores are 54.9% on Terminal-Bench 2.1 and 74.2% on SWE-bench Verified, which are release claims only. One user measured 114 to 132 tokens/s on an M5 Max with int4 MLX.
• DeepSeek DeepInfer is now open source. It is the production serving stack behind DeepSeek-V4.1-Flash. SemiAnalysis left it out of InferenceX because serving components are missing.
• Google TorchTPU is open-sourced. SGLang and vLLM plugins are available.
• EAGLE-4 reports a 3.42× speedup on SGLang on 4× B300 for Qwen3.8-27B at single-request concurrency. The gains shrink at higher concurrency, and the draft model is about 3 GB larger than DFlash2's.

Limits

The source is a newsletter that aggregates posts, not a primary advisory. Key gaps it names itself:

• exact affected and fixed versions, and deployment prerequisites
• what leaked and when
• scanning versus successful exploitation
• validated exposure counts
• the cause of each reported outage

The fetched text was cut off at 60,000 of 61,854 characters, during the Reddit recap. The catalog summary reports a 114,281-character page, so the fetched copy may differ from the version the summary was based on.