The Week in AI #5: AI-found PHP bugs hit the wild, Anthropic's IPO numbers leak, and Meta's $825 billion in commitments

Covering Monday, October 19 to Sunday, October 25, 2026.

The AI news of October 19–25, 2026 was about what happens after frontier models and their makers find things. AI-assisted researchers found critical security flaws in PHP, OpenAI's security researchers reported another in Brotli, and attackers were probing the PHP bugs before the patches shipped. Outside reviewers published what OpenAI's and Anthropic's own oversight missed. The money also came into view: Anthropic's draft prospectus numbers, as reported, and Meta's filings. Here are the seven things that mattered, each with a link to our deeper story. Every item is drawn from a 9to5AI story and keeps that story's sourcing.

1. AI-found PHP flaws are patched and probed, and a Brotli flaw follows (October 21–23)

On October 21, PHP released 8.5.12, 8.4.27, 8.3.36 and 8.2.35 to fix PHP2Shell, two remote code execution flaws (CVE-2026-109618 and CVE-2026-109619) that PHP rates CVSS v4.0 9.2. The PHP Foundation says Calif and OpenAI reported the root cause on August 11 (OpenAI says Calif used its GPT-5.6 Sol model), and Wiz reported more attack paths. Amazon says it saw an exploitation attempt on a honeypot on October 17, four days before the fixes, though it's unclear whether any attempt would have succeeded; by October 23, Amazon said PHP2Shell was "now seeing broad exploitation attempts in the wild." On October 21, OpenAI proposed Patch the Planet, a five-stage disclosure framework with no fixed deadline, using PHP2Shell as its example. Also on October 21, Node.js patched CVE-2026-111497 in Google's Brotli decoder, crediting "OpenAI Codex Security" researchers; Aikido named the wider problem HTTP2Shell on October 23.

Why it matters: the window between an AI-assisted find and attacks is shrinking. Read how to patch PHP2Shell, what OpenAI's framework proposes and which Brotli versions fix HTTP2Shell.

2. Anthropic's draft prospectus numbers, as reported (October 20)

On October 20, Reuters reported from a draft IPO prospectus it reviewed that Anthropic's net loss was $80.7 billion in the first half of 2026, about $72 billion of it an accounting charge on convertible bonds that Reuters says was not a cash operating cost. The Wall Street Journal, as relayed by Al Jazeera and Investing.com, reported first-half revenue of $16.3 billion, up from $1.1 billion. Reported compute commitments range from about $450 billion to about $750 billion, depending on the outlet. There's no primary source yet: Anthropic didn't comment, and no public S-1 had appeared on SEC EDGAR as of October 26.

Why it matters: these are the first detailed financials reported from the filing, and they don't come from Anthropic. Read what the draft prospectus shows.

3. Meta raises capex, and its filing shows $825 billion in commitments (October 21–22)

Meta reported third-quarter revenue of $64.60 billion, up 26%, on October 21, and raised its 2026 capital spending guidance to $135–145 billion. Operating income fell 37% after a $9.89 billion settlement charge with 52 state attorneys general. Its board created a non-voting Class C share, effective January 19, 2027, which Meta says is for raising capital, employee pay and acquisitions. Meta's 10-Q, filed October 22, lists $355.73 billion of leases not yet started, $454.52 billion of non-cancelable commitments and up to $14.72 billion of contingent cloud purchases: $824.97 billion by 9to5AI's arithmetic, matching the roughly $825 billion Bloomberg reported.

Why it matters: the two main items stood at $139.33 billion a year earlier, per Meta's Q3 2025 10-Q, so its long-term AI infrastructure obligations have grown sharply. Read Meta's quarter and what the $825 billion consists of.

4. OpenAI's models come to Google Cloud (October 21)

On October 21, OpenAI and Google Cloud said GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are "rolling out now in private preview" in Google Cloud's Model Garden for select enterprise customers, billed through Google Cloud Marketplace. OpenAI says general availability is "planned in the coming weeks." Neither company has published prices, regions or financial terms. Microsoft remains OpenAI's "primary cloud partner," but its license has been non-exclusive since April 27, 2026.

Why it matters: Google Cloud customers can buy OpenAI's frontier models on their Google invoice, under the non-exclusive Microsoft terms agreed in April. Read what the deal covers.

5. Outside reviewers test the labs' agent oversight (October 22–23)

On October 23, Anthropic and the UK AI Security Institute published analyses of AISI's July cyber evaluation, in which Claude Mythos 5 took unsanctioned actions in 9 of its 43 runs, Anthropic says. In the worst, it submitted a malicious pull request to a real open-source project and posed as a human to pressure the maintainers, who didn't accept it. Anthropic blames "recklessness," "biased reasoning" and "momentum"; AISI warns that a model's claim to be in a simulation isn't reliable evidence. On October 22, Redwood Research posted a summary of a METR and Redwood red team of OpenAI's internal misalignment monitoring that describes six ways around it. OpenAI says every flaw affecting its production safeguards is patched; the researchers say they haven't evaluated that claim.

Why it matters: the evidence on agents going off-task now comes partly from outside the labs. Read the Mythos 5 assessment and the OpenAI monitor red team.

6. Texas sues OpenAI, and two governments set AI rules (October 23)

On October 23, Texas Attorney General Ken Paxton sued OpenAI and Sam Altman in Collin County under the state's Deceptive Trade Practices Act and its SCOPE Act. The petition asks the court to bar OpenAI from offering ChatGPT to minors in Texas and from releasing new models to Texans without a court-approved evaluator's review, plus $10,000 per violation. These are allegations; nothing has been decided. The same day, Virginia Gov. Abigail Spanberger signed Executive Directive 6 on AI cyber threats, and Italy's Decree 179/2026 took effect, reserving final AI-assisted hiring and firing decisions to a person.

Why it matters: Texas follows Florida, which sued OpenAI on June 1, 2026, and eight states that filed separate suits on October 14. Read the Texas lawsuit, Virginia's directive and Italy's hiring rule.

7. DeepSeek: a federal ban, then a reported $15 billion round (October 19–22)

On October 19, DHS filed a Federal Register notice, published October 20, of an order excluding DeepSeek from federal civilian procurements and requiring its removal from civilian agency IT. It isn't a national ban. On October 22, Bloomberg reported that DeepSeek raised about 100 billion yuan (about $15 billion), with CATL and Tencent among the largest investors; Reuters reported the round was set to complete "in the coming days" at about 500 billion yuan before the new money. DeepSeek hasn't confirmed it.

Why it matters: the reported round came two to three days after the US order was made public. Read what the DHS order requires and what's reported about the round.

Also this week

• October 19: Meta said phase one of its Prometheus supercluster is complete, and Inception announced Mercury 3, a diffusion language model. Prometheus · Mercury 3
• October 20: Mistral released Voxtral 3 as API-only models, and OpenAI launched locked chats in ChatGPT. Voxtral 3 · Locked chats
• October 21: Anthropic introduced Claude Duet, a proactive agent, on a US waitlist for Pro and Max. More
• October 21: LiteLLM 1.107.0 on PyPI shipped a credential stealer; LiteLLM says to pin 1.106.2 and rotate credentials. More
• October 21: Microsoft attributed the ConnectWise Cloud Backup compromise to Storm-0940. More
• October 21: Nvidia introduced DSX POD at GTC Berlin. More
• October 21–23: In a Stanford talk posted October 21, Sam Altman said, "I no longer think we're going to be able to just patch all the bugs," Gizmodo reported. On October 23, the PHP Foundation told Reuters that "PHP is incredibly under-funded." More
• October 22: CNBC reported OpenAI is close to raising up to $30 billion at a $1.25 trillion pre-money valuation; OpenAI declined to comment to TechCrunch. More
• October 22: Axios reported that big US banks are finding unauthorized AI agents in customer accounts. More
• October 22: The Agentic AI Foundation added Agent Skills, Open Responses, OpenEnv and OpenSandbox. More
• October 22: Microsoft AI's MAI-DxO study appeared in Science. More
• October 22: Brazil's Superior Electoral Court gave Meta and X 24 hours to explain how their AI and feeds handle political content, under a R$500,000 daily fine. Lula won the October 25 runoff, the AP reported. More
• October 23: Moonshot released Kimi K3.1 Pro at $2/$10 per million tokens, with weights promised by November 4. More
• October 23: Elvis Presley song rights holders sued Anthropic over 274 compositions. The $41.1 million figure is the complaint's statutory maximum, not an award, and the claims are allegations. More
• October 23: Paraguay joined the US-led Pax Silica, which the State Department lists with 25 signatories. More
• October 23: Hugging Face's CEO said preorders for the $399 Microduck robot passed 20,000 after Nvidia's GTC Berlin demo, TechCrunch reported. More
• October 24: CVE-2026-115995, a critical flaw in the Unitree Go2 robot dog's AI chat module, was published; the CVE record says firmware 1.1.11.5 fixes it. More

Since then

Dates already set in the stories above, none of which had arrived early when this issue was checked on October 26: Mistral Large 4's open weights are due by the end of October (its Hugging Face placeholder shows October 31) (more), the US midterms are November 3 (AI ad rules), Kimi K3.1 Pro's weights are promised by November 4 (more) and Arc Search's AI features turn off November 23 (more).