{"slug": "ai-helped-find-two-flaws-in-php-code-behind-most-websites", "title": "AI helped find two flaws in PHP, code behind most websites", "summary": "Two critical PHP engine flaws rated 9.2 out of 10 for severity, affecting nearly every PHP release back to version 5.4, were found with the help of AI tools after going unnoticed for more than a decade, according to OpenAI, which said its GPT-5.6 Sol model helped security firm Calif find one flaw under an OpenAI-led effort called Patch the Planet. Wiz, which published the most detailed technical account, said it made its first discovery on 19 August and reported to PHP's security team on 24 August using its Wiz Atlas agent plus Google's Gemini 4 Argon and GPT-5.6 Sol, and that the flaws let an attacker run code on a server or read memory holding passwords, tokens, API keys and certificates without logging in. PHP moved its planned 22 October release forward to 21 October at 14:00 UTC after the CVE IDs leaked on 20 October, and governments and agencies urged sites to patch or go offline.", "body_md": "Two serious flaws in PHP, the programming language behind an [estimated 70% of websites](https://openai.com/index/coordinated-disclosure-of-critical-vulnerabilities/), were found with the help of AI tools after going unnoticed for [more than a decade](https://www.bbc.co.uk/news/articles/c3qjnwye0ep5o). OpenAI presents that as defenders getting there first. The days that followed complicated the picture: details leaked a day before the fix, governments and councils took sites offline, and Amazon says attackers were probing a decoy server before any fix existed.\n\n## Two flaws in the core of PHP\n\nThe 70% is OpenAI's estimate of how much of the web runs on PHP. It is not the share of sites that were vulnerable. [Wiz](https://www.wiz.io/blog/php2shell-php-rce), the cloud security company that published the most detailed technical account, says the flaws affect nearly every PHP release of the past decade, back to version 5.4. Several news outlets put the starting point at PHP 5, which is older.\n\nBoth flaws sit in PHP's core engine rather than in any one website's code, and each is rated [9.2 out of 10 for severity](https://www.helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619/). [HackerOne](https://www.hackerone.com/blog/critical-php-vulnerabilities-hackerone-response) describes one as corruption when PHP runs out of memory and the other as internal data that stays reachable after the memory holding it has been freed. Neither requires the attacker to log in.\n\nWiz says an attacker could run their own code on the server, or read the server's memory, which can hold passwords, authentication tokens, API keys and certificates. Tokens are the digital passes programs use to prove who they are. Wiz also says it demonstrated the attack with a single script against ready-made installations of WordPress, Drupal and Joomla, three systems for building websites, without logging in.\n\nOpenAI's president, Greg Brockman, [said](https://sea.mashable.com/tech/56312/openai-says-gpt-56-sol-helped-find-the-php-flaws-now-shaking-the-internet) the bug had existed for 14 years.\n\n## Who found them depends on who is telling it\n\nAI is part of every account, but the accounts differ on whose AI did what. OpenAI says a security firm called Calif found one critical flaw using GPT-5.6 Sol, OpenAI's model, as part of an OpenAI-led effort called Patch the Planet. It says Wiz independently found and reported both.\n\nWiz says it made its first discovery on 19 August and reported to PHP's security team on 24 August. It names three tools: its own AI security agent, Wiz Atlas, plus Google's Gemini 4 Argon and GPT-5.6 Sol. HackerOne's account has a different order, with Calif first and Wiz weeks later. Who was first, and whose model deserves the credit, is contested.\n\n## Two months of preparation, then a leak\n\nOpenAI says the fix was intricate and needed a major change to the PHP engine, so it led a coalition of maintainers, researchers and security providers for about two months to prepare fixes and protections before disclosure. That is the point of coordinated disclosure: the flaw stays secret until a fix is ready, so attackers get no head start. Wiz says PHP began developing its fix on 21 September.\n\nThe release had been planned for 22 October. After the flaws' CVE IDs leaked on 20 October, PHP moved it forward a day, to 21 October at 14:00 UTC. The same day, [Infosecurity Magazine](https://www.infosecurity-magazine.com/news/critical-php-vulnerabilities-leaked/) reports, details of both flaws leaked online before patches were out. PHP released its fixes on 21 October, at about 14:00 UTC.\n\n## Sites went offline instead of waiting\n\nThe BBC [reported](https://www.myjoyonline.com/websites-taken-down-worldwide-as-critical-cyber-security-update-issued/) that governments and agencies urged websites to deal with the flaws, and that organisations unable to update at once should consider taking their sites down. [ABC News](https://abcnews.com/Technology/urgent-security-fix-issued-code-vital-estimated-70/story?id=137447718) named the US cybersecurity agency CISA, the FBI's cyber division and the UK's National Cyber Security Centre among those giving that advice.\n\nSome organisations did. The BBC reported that the Finnish government, Edinburgh council and BBC programme pages went offline. In Scotland, [at least seven of 32 councils and six of 14 health boards](https://www.bbc.co.uk/news/articles/cwm23e6pe93mo) took websites offline as a precaution. The PHP Foundation, the nonprofit that supports the language, [said](https://www.wired.com/story/this-is-not-a-drill-2-php-vulnerabilities-are-setting-the-internet-on-fire/) there are no complete mitigations and that updating is the only safe choice.\n\n## Attacks came before the fix and after it\n\nAmazon Web Services says the first attempts came before release. Its chief information security officer, CJ Moses, [said](https://www.wired.com/story/hackers-exploited-the-php-vulnerabilities-days-before-they-were-patched/) a honeypot, a decoy server set up to attract attackers, was targeted on 17 October from an anonymised address while the flaws were still under embargo. He said it is unclear how the attacker got exploit information, and unclear whether the attempts would have worked on real systems.\n\nAfter release, Amazon Threat Intelligence [says](https://abcnews.com/Technology/urgent-security-fix-issued-code-vital-estimated-70/story?id=137447718) more proof-of-concept exploits appeared within hours and were used against its sensors almost immediately, and that the flaws are now being broadly exploited. Cloudflare's chief technology officer, Dane Knecht, had said before the fixes shipped that the company was seeing early attacks emerge.\n\n## Why the gap after a fix matters\n\nOpenAI says that, to its knowledge, AI found these flaws ahead of attackers. It also warns that the same ability can help attackers develop offensive tools and exploits quickly. Moses spelled out the second point. In the past, he said, most exploitation AWS saw came after someone published a working demonstration, because only a small number of capable actors built exploits and everyone else waited for them. He says that dependency is weakening: AI-assisted analysis makes it cheaper to turn a patch diff, the comparison between old and new code, or a security advisory into a working exploit, so more actors can act within hours or even minutes of disclosure. Wiz makes a related point, saying AI can help adapt an exploit to different PHP versions, configurations and memory layouts.\n\nThat puts the risk in the window between a fix going public and a site being updated. [The Decoder](https://the-decoder.com/openais-gpt-5-6-sol-and-gemini-4-argon-discovered-the-critical-php-vulnerability-affecting-millions-of-websites/) reports that OpenAI expects findings of this scale to increase sharply in the coming months, which would make that window come around more often.\n\n## What operators need to update\n\nThe fixed versions are [PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35](https://laravel.com/forge/docs/knowledge-base/php-security-release-october-2026). OpenAI released a [checker tool](https://github.com/openai/php2shell-checker) that site operators can use to test whether their PHP version is vulnerable. PHP 8.1 and earlier have reached end of life and get no official patch, although HackerOne notes minimal source backports for 7.4, 8.0 and 8.1. For a site on one of those versions, the official fix is to move to a supported release.", "url": "https://wpnews.pro/news/ai-helped-find-two-flaws-in-php-code-behind-most-websites", "canonical_source": "https://recul.ai/news/ai-helped-find-php-flaws-behind-most-websites/", "published_at": "2026-10-23 09:59:58+00:00", "updated_at": "2026-10-23 10:10:24.205666+00:00", "lang": "en", "topics": ["ai-safety", "artificial-intelligence", "generative-ai"], "entities": ["OpenAI", "PHP", "Wiz", "GPT-5.6 Sol", "Wiz Atlas", "Gemini 4 Argon", "HackerOne", "Greg Brockman"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ai-helped-find-two-flaws-in-php-code-behind-most-websites", "markdown": "https://wpnews.pro/news/ai-helped-find-two-flaws-in-php-code-behind-most-websites.md", "text": "https://wpnews.pro/news/ai-helped-find-two-flaws-in-php-code-behind-most-websites.txt", "jsonld": "https://wpnews.pro/news/ai-helped-find-two-flaws-in-php-code-behind-most-websites.jsonld"}}