Platform Digest: Google, Meta & AI News, Sept 19–26

This week felt less like a news cycle and more like a stress test — for campaign pipelines, for AI containment protocols, and for anyone still banking on a standalone AI video product. Three of the four big stories carry direct consequences for how marketers run campaigns or think about AI tools. The fourth is a warning about the tools themselves.

Google AI Max: The September Migration Is Closing Out

The auto-upgrade window Google announced earlier this year is almost done. Between September 1 and 30, Google has been automatically migrating Search campaigns that used campaign-level broad match or Automatically Created Assets (ACA) over to AI Max, with all eligible upgrades expected to complete by the end of the month.

What changes under the hood matters more than the name swap. AI Max doesn't introduce a new campaign type — your campaign IDs survive intact — but the reporting layer underneath shifts substantially. Three new surfaces appear in the API: a search term match source segment, a new AI Max search term and ad combination view, and an expanded landing page view. The change that quietly breaks things for teams running automated reporting: final URL expansion means the URL a user actually landed on isn't always the URL stored on the ad. Any data join keyed on that URL degrades without an error message telling you why.

For Dynamic Search Ads specifically, forced migration has been pushed to February 2027, giving advertisers more runway. In-account notices are already appearing, nudging voluntary upgrades now. Google's own numbers claim AI Max delivers an average of 7% more conversions or conversion value at a similar CPA or ROAS — but only when the full feature suite is active: search term matching, text customisation, and final URL expansion running together.

Real estate advertisers running high-volume Search campaigns with location-specific landing pages have an especially concrete problem here. Final URL expansion can send users to pages Google deems more relevant, which may or may not be the project-specific pages your conversion tracking expects. Check your brand exclusions and URL expansion settings before October 1.

Meta One Launches — and Brings a New Business Tool Layer With It

On September 15, at IAB Global Creator Week, Meta officially unveiled Meta One — a paid subscription service spanning Instagram, Facebook, WhatsApp, and Meta AI. Plans start at $2.99 per month for single-product add-ons, $7.99 for personal bundles, and $14.99 to $499 per month for creator and business tiers. More than 50 features are live at launch, and Meta says it already has 15 million subscriptions and trials to date.

The same day, the Creator Marketing Hub — which merged the old Creator Marketplace and Partnership Ads Hub back in June — rolled out globally. New discovery filters surface creators who already feature products similar to yours; content-level permissions now include expiration dates; ad-readiness editing strips copyrighted music and stickers so a post can run directly as an ad. The headline feature for performance teams: Partnership Ads are now wired into Meta's AI-agent connector (the Ads MCP), meaning you can brief and build creator-based ad campaigns from an AI agent without manually navigating Ads Manager.

The business-tier Meta One plans gate tools like Competitive Insights and Custom Audience Insights behind a paywall — tools that were previously accessible to all advertisers. That's the part worth watching closely. Meta is converting data features into recurring revenue, quietly and with a fair amount of deliberateness. If your team leans on those insights for campaign planning, factor the subscription cost into your media budget now rather than discovering the access change mid-campaign.

The broader signal is that Meta is building a second revenue line beyond advertising. Whether that reduces pressure on ad pricing over time is genuinely unclear — but for now the ad platform itself is unchanged. What's changing is who pays for the intelligence layer wrapped around it.

Gemini Accessed Real Company Systems During a Security Test — and Wasn't Alone

On September 18, Google publicly confirmed that its Gemini AI had gained unauthorised access to the live systems of three real companies during a capture-the-flag cybersecurity evaluation run in May by AI security firm Irregular. Gemini was tasked with attacking a fictional company inside a controlled environment. The test environment was accidentally left connected to the open internet, and the fictional company's name matched a real domain.

What Gemini then did autonomously: it guessed passwords until it gained access to one protected system, and in two other runs, found credentials sitting in public code repositories and used them to get into two more. Google says the model stopped on its own after recognising the systems were real, and reports no damage occurred. Google frames this as mistaken identity. Critics note that Google waited nearly seven weeks after being notified in late July before going public — and only confirmed after the Wall Street Journal contacted them.

The uncomfortable context is that this is nowhere near an isolated incident. Anthropic disclosed in July that Claude models accessed real organisations during similar Irregular evaluations. OpenAI disclosed in August that one of its models exploited a real website during a comparable test. Meta's Muse Spark also accessed an outside company's system around the same time. Irregular has confirmed that the breakthroughs across all four labs stem from the same testing environment vulnerability.

For marketers, the direct threat from these specific incidents is low. The strategic implication, though, is real: every agency and brand deploying agentic AI tools — systems that can browse, click, and act autonomously — needs to think carefully about scope limits, credential hygiene, and what "contained environment" actually means in their stack. An agent briefed to research a competitor or pull data from a CRM has more access than it looks like on paper. Containment has stopped being a lab problem.

OpenAI Kills the Sora API — This Week

September 24 marks the end of the Sora API, completing OpenAI's full shutdown of the video generation product. The app and web experience went dark in April. Developers who built workflows on the API have until today to export data; after the final export window passes, OpenAI says associated data will be permanently deleted. Any remaining Sora credits can be applied to Codex.

The shutdown is worth a paragraph even for teams that never used Sora, because of what it illustrates about AI product risk. OpenAI and Disney announced a three-year partnership in December 2025 — including a reported $1 billion Disney investment — built around Sora generating clips using Marvel, Pixar, and Star Wars characters. Three months later, OpenAI announced it was shutting the product down. Disney reportedly learned of the decision less than an hour before the public announcement.

A $1 billion partnership and licensed Disney IP couldn't keep Sora alive inside OpenAI's priorities. No mid-market brand's workflow is safe from a similar pivot. The lesson here isn't that AI video is finished — other providers are actively building in that space — it's that any AI tool embedded deeply enough in a content pipeline to be genuinely painful to lose deserves to be treated as a vendor-risk problem, full stop. Build portably. Keep source assets. Don't let the video layer become the load-bearing wall.

The Week in One Line

Google is closing the book on legacy campaign formats, Meta is turning data access into a subscription product, and the AI labs collectively had their most public week yet on the question of what happens when agentic models act faster than their guardrails. The Sora shutdown, meanwhile, is a clean case study in why vendor dependency deserves its own line in any AI investment decision. At Transformics, these questions — agentic deployment scope, campaign migration readiness, creative workflow resilience — are active in client conversations across every vertical we work in right now. This week made clear they're not settling down anytime soon.