AI
“Security firms this week revealed they used AI to discover severe, previously undiscovered security holes in PHP, open source software that underpins much of the internet including banking, airlines and government services, that had gone undetected for more than a decade.”
Plain restatementIn the week of October 19 to 23, 2026, security companies publicly disclosed that AI tools assisted them in finding critical vulnerabilities in the PHP programming language, and those vulnerabilities had been present and unaddressed for over ten years.
This claim is mostly accurate. On October 21, 2026, the PHP project released emergency patches for two critical flaws, nicknamed PHP2Shell, that let an attacker run code on a web server without logging in, and both were rated 9.2 out of 10 for severity. The vulnerable code goes back to a PHP version released in 2012, so "more than a decade" holds, and security firms Calif and Wiz, working with OpenAI, did use AI models in the research. Banks, airlines and government sites in several countries did take services offline as a precaution. The main wrinkle is the phrase "previously undiscovered." The PHP Foundation has said one of its own developers spotted signs of the underlying problem during a 2024 audit but could not find a way to exploit it and the fix stalled, so what the AI work mainly changed was proving the flaw could actually be attacked. It is also worth knowing that the strongest statements about how much the AI mattered come from OpenAI and Wiz, who have a commercial interest in that framing, and that one of the discovering firms gives a more modest account crediting a human researcher's hunch.
[drifted from the evidence:] Security firms this week [drifted from the evidence:] revealed they used AI to [drifted from the evidence:] discover severe, previously undiscovered security [drifted from the evidence:] holes in PHP, open source software that [drifted from the evidence:] underpins much of the [drifted from the evidence:] internet including banking, airlines and [drifted from the evidence:] government services, that had [drifted from the evidence:] gone undetected for [drifted from the evidence:] more than a decade.
[added by the neutral restatement:] In the week [added by the neutral restatement:] of October 19 to [added by the neutral restatement:] 23, 2026, security [added by the neutral restatement:] companies publicly disclosed that [added by the neutral restatement:] AI tools assisted them in finding critical vulnerabilities in the [added by the neutral restatement:] PHP programming language, and [added by the neutral restatement:] those vulnerabilities had [added by the neutral restatement:] been present and unaddressed for [added by the neutral restatement:] over ten years.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- Two critical PHP core vulnerabilities, CVE-2026-109618 and CVE-2026-109619, were publicly disclosed and patched on October 21, 2026, both rated CVSS v4.0 9.2 by the PHP security team. The claim's "this week" is correct relative to the post's October 23 publication.
- PHP is open source and extremely widely deployed. Multiple sources put its share at roughly 70% of websites, and the flaws reached banking, airline and government systems in practice, with Japanese and European examples documented.
- The vulnerable code is genuinely old. Wiz states every PHP version back to 5.4, released in 2012, is affected by at least one of the vulnerabilities, which supports "more than a decade."
- AI tools were genuinely used in the research. OpenAI's own page names GPT-5.6 Sol as used by Calif, and Wiz names Gemini 4 Argon, GPT-5.6 Sol and Wiz Atlas.
- The discovering organizations include security firms. Calif and Wiz are security companies; Wiz is owned by Google.
- The flaws were not previously fixed, and no working exploit path for them was publicly known before 2026.
- On the post's other statements: the 22 million figure traces to a real primary source, with Unit 42 reporting over 22.6 million exposed instances potentially vulnerable based on Cortex Xpanse telemetry. UK outlets reported 32 UK government services offline due to the flaw, slightly more than the post's "30." Wes Streeting was Defence Secretary and did urge the public to stockpile days' worth of food and water, citing increasingly sophisticated cyber attacks, after 32 government services were taken offline. The Anthropic quote is accurate: Anthropic deputy CISO Jason Clinton wrote that as bad as things have gotten in the past four days, they are going to get a lot worse before they get better.
- The claim says the holes were "previously undiscovered." For one of the two CVEs that is imprecise. The PHP Foundation's own account states a core developer found signs of the out-of-memory flaw behind CVE-2026-109618 during PHP's 2024 external audit, but no route to code execution was found and the fix stalled. The accurate framing is that the flaw was partially seen and deprioritized, not wholly unknown. This matters because it shifts part of the story from "AI saw what nobody saw" toward "AI proved exploitable what maintainers had seen and could not act on."
- The claim compresses "AI was used to find the flaws" without the discoverers' own narrower framing. Wiz's stated position is that the code was long present and what AI changed was the ability to turn the weakness into a working exploit. The claim's wording is consistent with Wiz's public headline but omits the distinction between surfacing code and demonstrating exploitability, which is the actual technical contribution described.
- The strongest assertions about AI's role, including that frontier AI found the flaws before attackers, come from OpenAI's own promotional post for its Patch the Planet initiative and from Wiz, a security vendor and Google subsidiary. These are interested parties on precisely the question of how much AI mattered. One of the discoverers, Calif, publicly describes the initial discovery as driven by a human researcher's hypothesis with no special harness, which is a materially more modest account from inside the same disclosure.
- Misattribution, minor: the claim attributes the work to "security firms." Calif and Wiz are security firms, but OpenAI is an AI developer and was an independent reporter of one root cause, not a security firm. This is a small imprecision rather than a reversal.
- On the post's wider statements, one figure did not check out as attributed: the post says Palo Alto Networks estimates roughly a third of the 22 million involves legacy software that cannot be patched at all. Unit 42's published brief gives the 22.6 million exposure figure but no one-third legacy breakdown was located in it, and other published scans of version-disclosing PHP services put the end-of-life share considerably higher, with one scan of 2,220,231 version-disclosing services finding nearly 73% on unsupported branches and Qualys reporting 110 million of 156 million installations with a known version, or 71%, on end-of-life or end-of-support releases. Note that these scans measure different populations than Unit 42's exposed-instance count and are not directly comparable.
- How much the AI models contributed relative to human researchers remains genuinely disputed between the discovering parties themselves, with OpenAI asserting frontier AI was key and Calif crediting a human hypothesis and no special tooling in the initial discovery. No independent third-party reconstruction of the research process has been published.
- Whether attackers had independently found the flaws before disclosure is not settled. OpenAI's "to our knowledge" phrasing is itself hedged, and separate reporting of pre-disclosure exploitation attempts beginning around October 17 complicates the timeline.
- The exact count of systems actually vulnerable, as distinct from exposed and potentially vulnerable, is not established. Unit 42's own wording is "potentially vulnerable," and multiple scanners produce widely divergent totals because they measure different things.
- Whether the Defence Secretary's stockpiling advice was prompted by the PHP incident specifically, or whether the PHP incident was cited within a pre-planned broader preparedness campaign, is not fully resolved by the sources found. Reporting indicates the "Be Ready" campaign is a Cabinet Office effort framed around hostile state actors and terrorists generally, with the PHP episode as immediate context.
The underlying event is real, recent, and extensively documented by primary artifacts. Two critical vulnerabilities in PHP core were disclosed on October 21, 2026. The PHP project released PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35 to fix two vulnerabilities, both rated CVSS v4.0 9.2, that can allow unauthenticated remote code execution, after the PHP security team published an advance notice on October 20. The Canadian Centre for Cyber Security issued an alert on October 21 describing CVE-2026-109618 and CVE-2026-109619 as use-after-free vulnerabilities affecting multiple PHP versions, and stated that all PHP installations serving web requests, including PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP and LiteSpeed, are at risk. On the age of the flaws, the discoverer's own artifact is explicit. Wiz stated that every PHP version released in the last decade is affected by at least one of these vulnerabilities, going back to version 5.4, that attackers can exploit them remotely, and that the findings rank among the most severe vulnerabilities discovered in PHP, and arguably in software history. On the AI role, Wiz described it narrowly. Wiz wrote that the underlying flaws had been present for decades, but demonstrating their exploitability required an exceptionally complex chain of steps at the edge of what human researchers could practically achieve, and that frontier AI models substantially reduced that barrier. Fortune quoted Wiz saying the vulnerable code had been there for more than a decade and that what AI changed was the ability to understand the weaknesses and turn them into a working exploit. OpenAI's own page makes a parallel but distinct claim. OpenAI said that given the severity of the flaws and PHP's widespread use it considers this one of the most impactful security issues in recent memory, and that to its knowledge frontier AI found these vulnerabilities before attackers did, opening a Defender's Window for maintainers and security providers. The attribution on OpenAI's page is specific: Calif identified a critical flaw using GPT-5.6 Sol as part of the Patch the Planet initiative, with Wiz also independently discovering and reporting both vulnerabilities. A material qualifier sits in the PHP Foundation's own post-incident account. The PHP Foundation said a core developer first found signs of the out-of-memory flaw behind CVE-2026-109618 during PHP's 2024 external audit, but a working attack was only theorized and no concrete route to code execution was found then; the 2026 reports from Calif, OpenAI and Wiz, which used AI tools, established the working exploit that made the flaw's full impact clear and moved the fix to the top of the queue. Fortune reported Brockman's account that the bug had been present in PHP for 14 years and that a limited version of it had been discovered in 2024 and was not considered significant at the time. The AI contribution is also partly contested among the discoverers themselves. One technical writeup records that how much the models mattered is contested: OpenAI said frontier AI was key, while Calif's own write-up says no special harness or non-public model was used in the initial discovery and credits a theory about where PHP's old bugs hide. The banking, airline and government disruption described in the claim is documented. Japanese airline ANA, railway operator JR West and the country's largest bank MUFG were among companies that reported some webpages and services temporarily suspended, and government websites in France, Denmark, Finland and the Netherlands were affected.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Were these PHP security flaws really unknown until AI found them?
Not entirely. The PHP Foundation says a developer spotted signs of one of the two flaws during a 2024 audit but could not find a way to exploit it, so the fix stalled. The 2026 AI-assisted research mainly proved the flaw could actually be attacked, rather than discovering it from scratch.
How old were the vulnerabilities?
Very old. Wiz says every PHP version released since 5.4 in 2012 is affected by at least one of the two flaws, which supports the claim that they went undetected for more than a decade.
What did AI actually do in this research?
Security firms Calif and Wiz, along with OpenAI, used AI models including GPT-5.6 Sol, Gemini 4 Argon and Wiz Atlas to help turn the known weaknesses into working exploits. Wiz describes the AI's role as reducing the complexity barrier to proving exploitability, not necessarily spotting the bugs first.
Did AI deserve all the credit, or is that disputed?
It is disputed among the discoverers themselves. OpenAI and Wiz, who have a commercial interest in promoting AI's role, give the strongest accounts of AI's importance, while Calif's own writeup says no special AI harness was used in the initial discovery and credits a human researcher's theory instead.
Did the vulnerabilities actually affect real banks, airlines and governments?
Yes. Japanese airline ANA, railway operator JR West, bank MUFG, and government websites in France, Denmark, Finland and the Netherlands reported services temporarily affected or taken offline as a precaution.