A machine in the headline, people in the body

Three headlines give the discovery to a machine. PBS, running Associated Press copy, opens with “AI discovered”Headline asserts, body hedgesThe mother pattern behind scare quotes, modal deletion, and the commitment-free statement: a headline states a claim as settled fact while the article underneath it quietly hedges, denies, or contradicts it. a software vulnerability [7]. RFI writes “Une IA découvre” (An AI discovers) a flaw in PHP [8]. France 24 has thousands of sites taken offline over a flaw “repérée par l’IA” (spotted by AI) [5]. Each body says something more modest. PBS says “Artificial intelligence agents helped uncover the vulnerabilities” [7]. France 24 says the flaws were “trouvées par l’entreprise de sécurité informatique Calif grâce à l’un de ses modèles les plus récents” (found by the IT security company Calif thanks to one of its most recent models) [5]. RFI says AI models “ont été utilisées par la société de cybersécurité Wiz et d’autres groupes” (were used by the cybersecurity company Wiz and other groups) [8].

The mother pattern behind scare quotes, modal deletion, and the commitment-free statement: a headline states a claim as settled fact while the article underneath it quietly hedges, denies, or contradicts it.

This is headline asserts, body hedges. Most readers only see the headline, so what they take away is that software found the flaw by itself. The body describes something else: named firms using a tool. This matters because the AI-as-discoverer version is also the version the AI’s vendor is promoting. To spot it, compare the headline’s grammatical subject with the subject of the first sentence that says who did the work.

The Guardian is the one outlet that puts this claim in doubt. It reports that OpenAI “claimed ‘frontier AI’ had been key”, and then adds: “The researcher who found the weakness, however, said they did not use models that were not available to the general public” [1]. None of the other outlets mentions this. They also disagree about which models were involved. France Info names OpenAI’s GPT‑5.6 Sol used by Calif and says the foundation had access to OpenAI and Anthropic tools [3]. RFI names “OpenAI et Google” [8]. France 24 says Wiz did not say which models it used [5]. From this coverage, a reader cannot tell what part the AI played.

“Before the experts”: a claim the outlets’ own facts contradict

RFI’s standfirst goes further. It says the flaw “a été repérée par des intelligences artificielles, avant que les experts ne la détectent” (was spotted by artificial intelligences before the experts detected it) [8]. Its own body says experts at Wiz ran the tools [8]. Other outlets report a longer history. France Info, drawing on the PHP Foundation’s site, says the first signs of one vulnerability were found in 2024 [3]. CBS quotes the reporter who broke the story: “People knew something weird was going on” [2]. BBC News mentions in passing that the flaw was “theorised by the PHP Foundation in 2024”Buried ledeA buried lede is the fact that should have organized the whole story, relegated instead to the final sentence or hidden inside someone else's quote. [4], in a single line where it raises no questions.

A buried lede is the fact that should have organized the whole story, relegated instead to the final sentence or hidden inside someone else's quote.

BBC News’s explainer opens in its own voice: “It took artificial intelligence (AI) to bring them to light” [6]. Further down, it quotes a researcher who puts it more carefully: “The underlying bugs weren’t new; what changed was our ability to find them” [6].

Firmer than the source

Some sentences state the threat or the outcome more firmly than the sources they rely on. AP reports OpenAI’s own claim with a hedge: “It said AI appeared to find the vulnerabilities before any attackers did” [7]. CBS’s first sentence has researchers finding the flaw “before hackers could exploit it”Certainty upgradeA certainty upgrade restates a source's claim in the outlet's own voice more firmly than the source made it: a hedge, a limit or a caveat in the original disappears in the retelling. [2], stated as fact and attributed to no one. That is a certainty upgrade: the vendor’s “appeared to” has disappeared, and the reader is told as settled what the company only suggested. The test is to look for the hedge in the original claim and check whether it survives the retelling.

A certainty upgrade restates a source's claim in the outlet's own voice more firmly than the source made it: a hedge, a limit or a caveat in the original disappears in the retelling.

The same move happens with the risk. The PHP Foundation said that “under the right conditions” an unauthenticated attacker could run code remotely [4]. The BBC explainer turns this into “everything built on PHP was left exposed” [6]. Its own sister article [4] keeps the condition.

The leak with no one leaking

The rush to take sites offline came after the disclosure plan broke down. RFI tells this in one sentence: “l’information a fuité”Agentless passiveThe agentless passive is a sentence built so the action happens with no one doing it. (the information leaked) [8]. BBC News writes that details “leaked online” this week [6], the same construction. This is the agentless passive: the leak just happens. The reader never asks who leaked it, even though France Info names them. According to the PHP Foundation, Calif published a proof-of-concept video “par erreur” (by mistake), and Chainguard, which had received the patch early, “a rompu l’embargo de manière unilatérale” (unilaterally broke the embargo) [3]. France Info also gives Chainguard’s side: its chief executive says he followed coordinated-disclosure guidelines because exploitation attempts had already begun [3]. So there is a real dispute, with named parties. In most of the coverage it shows up as weather.

The agentless passive is a sentence built so the action happens with no one doing it.

The maintainers’ admission, deep in the text

The most revealing sentence in the whole coverage is in the last third of France Info’s piece. The foundation writes: “Nous n’allons pas le nier : ce retard est un échec structurel, lié à la fois au processus décisionnel de l’équipe de sécurité et à son manque de ressources” (We won’t deny it: this delay is a structural failure, linked both to the security team’s decision-making and to its lack of resources) [3]. France Info adds that the non-profit “manquait de moyens humains” (lacked staff) [3]. Software that runs about 70% of the web was being maintained by an understaffed team that had seen the warning signs two years earlier. That is a buried lede. It could have framed the whole story; instead it comes after the patching advice. The Guardian gets closest, ending on Michał Woźniak’s view that “This is an environmental problem” in an industry with “an incentive to deploy a thing and then never update it” [1]. No outlet asks who funds the people who maintain PHP.

The vendor’s sentence everywhere

One line appears again and again: OpenAI’s description of the flaws as “one of the most impactful security issues in recent memory” [6] [7], in French at France 24 [5] and at RFI as « l’une des vulnérabilités les plus marquantes de ces dernières années » (one of the most significant vulnerabilities in recent years) [8]. Each outlet attributes it. Even so, the company whose model is being credited is the one assessing how big the event is, and no outlet compares that assessment with the foundation’s own 9.2 out of 10 severity score [3] or with the agencies’ wording.

What a reader would need instead

5 checks to run on the next story like this one.

  1. The primary record. The PHP Foundation’s advisory on php.net (which CBS mentions [2]) and the CVE identifiers France Info prints (CVE-2026-109618 and CVE-2026-109619 [3]). These say what the flaws do and under what conditions.

    Gap in the coverage

  2. A dated disclosure timeline. When the foundation first saw signs in 2024, when Calif and Wiz reported the flaws, who leaked what and when, and where the embargo dispute stands.

    Agentless passiveBuried lede

  3. A clear account of the AI’s role. Which model each team used, whether it was publicly available (the point the Guardian raises [1]), and what the people did.

    Headline asserts, body hedges

  4. The funding question. Who pays the foundation’s security team, and what it says it has changed.

    Buried lede

  5. Ongoing tracking. The national agencies already following exploitation and patching: ANSSI, the NCSC, CISA and ENISA [3] [4] [7]. They, not a vendor’s blog, are the place to check over the coming weeks whether any of the “millions” of attempts Cloudflare counted [7] got through.

    Gap in the coverage