Startups

AI helped find critical PHP flaws, and a leak forced a rushed fix

Wiz and Calif found the PHP2Shell bugs with help from AI models, and PHP patched a day early after a leak. Now PHP has flagged a further flaw the patch doesn't fix, and officials say to take systems offline if they can't patch.

By Robert Harrison · Oct 23, 2026, 10:30 PM EDT · Updated Oct 24, 2026, 12:25 AM EDT

Why it matters: A new PHP flaw isn't fixed by the patch, officials say take systems offline if they can't patch, and AI that finds such bugs can help attackers too, raising the stakes for AI security firms.

For information only, not investment advice.

The PHP patch that governments and companies rushed to install this week doesn't cover everything. On Oct. 23, PHP published an advisory on a further remote code execution vector, CVE-2026-115560, in its session upload progress feature, and Canada's Cyber Centre says it is not addressed by the Oct. 21 security updates. PHP says the issue "can only occur under an unusual combination of configuration and application code" and will be handled in the regular release cycle. There's no fixed version yet; the workaround is disabling the session.upload_progress.enabled setting.

Officials are turning up the volume. CISA updated its alert on Oct. 23 and said organizations that can't update immediately, or that run PHP 8.1 or earlier, "should take PHP-based systems offline until they can be patched or updated." The FBI said the same in a public service announcement, and the UK's National Cyber Security Centre said to consider it too.

So, has anyone actually been hacked?

Depends on who you ask. CISA says it has not observed successful exploitation in the wild, and Canada's Cyber Centre has identified no active exploitation, "but that can change quickly." Per the Guardian, the NCSC sent 50,000 alerts and had no reports of successful hacks. Sophos walked back an earlier assessment, saying its telemetry shows scanning and "a small number of failed exploitation attempts." Calif's Ryan Austin told the Guardian, "At the moment it doesn’t look like exploitation has taken off or been widespread."

Attempts are another story. Cloudflare CEO Matthew Prince wrote on X, as the Guardian reported, that attempts hit "hundreds of thousands within two hours, and millions now." Singapore's Cyber Security Agency reports widespread exploitation attempts. Amazon CISO CJ Moses told the BBC the Oct. 21 attacks varied enough to suggest multiple actors, and security firm Hadrian warned that day that the flaws were being exploited "in the wild at scale," per TechTarget. Lots of knocking; none of these agencies has confirmed that anyone got through.

Governments go dark to patch

The UK took 32 government services offline, cybersecurity minister Liz Lloyd told parliament, per the Guardian, and Axios reported dozens of websites in the U.S., U.K. and elsewhere briefly went down to patch. South Korea activated emergency measures, and Japan's prime minister encouraged businesses to patch or shut down their sites, the Guardian reports. Palo Alto Networks says its PAN-OS firewalls are affected, with fixes still to come, though it is not aware of any malicious exploitation against its products.

CNBC says the moment is being described as "Black October." Former CISA director Chris Krebs told CNBC that "3% of all critical vulnerabilities ever discovered have been found in the last month," his figure, which hasn't been independently verified.

Who found it, and with what

The original flaws, which Wiz calls PHP2Shell, are CVE-2026-109618 and CVE-2026-109619, and both carry a CVSS score of 9.2, per TechTarget. Wiz says every PHP version from the last decade is affected by at least one. OpenAI says Calif found a critical flaw using its GPT-5.6 Sol model and that Wiz, which CNBC and TechCrunch describe as Google-owned, independently found both. Wiz dates its first find to Aug. 19, using Google's Gemini 4 Argon and GPT-5.6 Sol; TechTarget, citing OpenAI's Clint Gibler, says Calif found the bug class on Aug. 11.

The fix was meant to be quiet. Per the PHP Foundation, via TechTarget, the CVE IDs leaked onto forums Oct. 20. TechTarget reports Calif then inadvertently posted a proof-of-concept video and Chainguard shipped images with the fix more than six hours early; Chainguard's CEO disputed the Foundation's account of why. PHP released the patch on Oct. 21, a day ahead of plan.

The pitch and the catch

OpenAI says, "To our knowledge, frontier AI found these vulnerabilities before attackers did," and calls this "one of the most impactful security issues in recent memory," its own characterization. It concedes the same capabilities "can also help attackers rapidly develop offensive tools and exploits," committed $1 billion in subsidized access to its Daybreak security offering, and expects similar findings "to increase sharply in the coming months."

Separately, and not a PHP bug, OpenAI said in a post on X that an internal red-teaming exercise on its own research environments turned up a remote code execution flaw in Brotli, a compression library included in Node.js, Bun and Envoy. For AI security companies, that's the pitch and the problem in one week: the models find the bugs, and the same models can help weaponize them.

Tags: ai · big-tech · regulation