PHP2Shell anatomy: two RCE bugs AI found, the Defender's Window, and the disclosure that leaked early

机器之心 · wechat · 2026-10-22

A deeper technical rundown of the PHP2Shell disclosure. GPT-5.6Sol, auditing PHP under OpenAI's Patch the Planet program, surfaced CVE-2026-109618 (memory management) and CVE-2026-109619 (use-after-free in error handling), both yielding unauthenticated RCE on Apache+modphp; affected versions span PHP 5.4 through 8.5. Wiz independently found the same flaws using its own WizAtlas agent plus Gemini 4 Argon and GPT-5.6Sol. OpenAI frames the episode around a "Defender's Window": AI finds internet-scale bugs before attackers do.

Key facts:

A 2024 external audit had already hinted at the OOM issue, but PHP maintainers stalled over backward-compatibility and DoS tradeoffs—what the PHP Foundation calls a "structural failure"—until AI found an actual exploit path
The five-stage disclosure plan collapsed on Oct 20: CVE numbers leaked, Calif accidentally posted a PoC video, and Chainguard shipped patched binaries early; Cloudflare observed early exploitation and honeypots caught live attacks
CISA urged immediate patching, the NSA suggested taking unpatchable sites offline, and Japanese games like Umamusume and Granblue Fantasy went down
OpenAI predicts AI-discovered vulnerabilities will sharply increase, pointing to memory-safe languages, formal verification, and sandboxing as long-term defenses

Original post →

PHP2Shell anatomy: two RCE bugs AI found, the Defender's Window, and the disclosure that leaked early

More from Models