Twenty stories that moved technology in September 2026

In brief

September 2026 in one sentence: AI got cheaper to buy by the token and dearer to run by the month. Frontier model prices fell, but agents moved onto usage meters, GPU rents rose, and regulators began handing data centres the bill for the grid. The proven security damage came through the network appliances organisations expose to the internet, not through the autonomous AI attacks that drew the headlines. Courts and regulators, not legislatures, set most of the month's rules. And the phones in people's pockets got more expensive, with their headline AI feature rationed.

Here are twenty stories that shaped the month, grouped by desk, each with its date and source. Where a story rests on one outlet, on anonymous sources or on a company's own claim, or where it is contested, the item says so.

AI

1. Anthropic cut frontier prices by 20% with Claude Opus 5.5 (22 September). The model costs $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Anthropic says it matches its larger Fable 5.1 "on most work", and its benchmark results are its own (Anthropic; TechCrunch). OpenAI added GPT-6 Sol and Luna the same day. OpenAI's own release notes confirm the models, but a reported 50% price cut appears only on an aggregator. ByteNib's analysis: frontier AI got cheaper, but budget for the meter.

2. Every large US lab now gates its strongest cyber model (1–3 September). Anthropic limited Mythos 5.1 to vetted US organisations, and Google opened Gemini 3.8 Flash Cyber only to "trusted defenders". OpenAI's GPT-6 Astra became its first model rated "Critical" for cybersecurity and ships switched off by default in enterprise workspaces (Anthropic; Google; CSO Online). One budgeting trap came with them: Gemini 3.8 Flash's introductory price doubles on 1 January 2027.

3. Washington accused six Chinese AI companies of "industrial-scale distillation" (8 September). The NSA, CISA and FBI advisory AA26-251A names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. It alleges campaigns against Claude, GPT, Gemini and Grok, made by SpaceXAI (formerly xAI) (CISA). China's Commerce Ministry called the accusation "baseless" the next day. On 10 September Anthropic alleged that Moonshot had relayed almost 300,000 real Kimi customer requests to Claude. On 22 September, reports that all trace back to The Information said China's internet regulator was questioning DeepSeek and Moonshot. The advisory is guidance, not charges or sanctions. See what it means for AI teams.

4. AI's legal exposure widened on two fronts. British Columbia sued OpenAI and Sam Altman in federal court in San Francisco. It alleges negligence after ChatGPT's safety team flagged the Tumbler Ridge shooter's conversations but did not alert police. The suit was filed on 21 September and widely reported on the 22nd, and summaries disagree on the filing day (Al Jazeera; CBC). On 17 September, unsealed filings in The New York Times' case against Microsoft and OpenAI quoted a Microsoft director calling AI scraping "the largest theft of labor in human history" (TechCrunch). The first US jury trial on AI training, Andersen v. Stability AI, has been pushed to April 2027. It did not take place in September.

5. The money kept coming (8–16 September). Mistral raised €3 billion at a valuation above €21 billion in a round led by Samsung (Mistral). Cognition, the maker of the Devin coding agent, raised more than $2 billion at $48 billion (Bloomberg). Bloomberg reported on 15 September, citing anonymous sources, that OpenAI is weighing a round valuing it at more than $1.2 trillion. CNBC said formal fundraising had not begun, and Forbes cited up to $1.5 trillion, so treat the figure as contested.

Security

6. Microsoft's September Patch Tuesday was the largest on record, by the analysts' count (8 September). It fixed 972 CVEs by CrowdStrike's and ZDI's count, with trackers ranging from 964 to 974, and CrowdStrike, Tenable, ZDI and BleepingComputer all call it a record. Only two of the flaws were already being exploited, both local privilege escalations. One of them, in Windows' ALPC component, was chained behind two Chrome zero-days in an espionage kit that Proofpoint attributes to the China-linked APT31 (Tenable; CrowdStrike; BleepingComputer). Analysts link the volume to AI-assisted bug discovery rather than to more attacks.

7. An attacker used hundreds of AI agents against PaperCut print servers. A likely Russian-speaking attacker used hundreds of AI agents, an OpenAI Codex harness running a DeepSeek model, to exploit two PaperCut print-server flaws. According to threat-intelligence firm GreyNoise, which reported the campaign on 9 September, the attacker compromised at least 440 servers at 395 organisations in 48 countries from 31 August. It harvested credentials at 280 victims and reached domain-administrator level at 12 (GreyNoise). A human set the targets and paced the follow-up, and "machine speed" is not GreyNoise's phrase. The AI details rest on GreyNoise alone. See what is proven about AI agents in attacks.

8. Exposed appliances were exploited, on three-day patch deadlines (14–22 September). Cisco disclosed exploitation of its Secure Email Gateway on 14 September. On 16 September it patched Identity Services Engine CVE-2026-76460, an authentication bypass rated CVSS 10 with no workaround. On 22 September CISA added exploited flaws in two Check Point products, F5 BIG-IP APM and Arista VeloCloud Orchestrator to its catalogue, each with a three-day federal deadline (Cisco; CISA KEV). The security firm Truesec reports that one Check Point flaw was attacked from 23 July, before any patch existed. See September's exploited edge.

9. Web platforms and supply chains were hit within hours. WordPress 7.1.2 fixed an unauthenticated remote-code-execution flaw on 22 September, and Patchstack saw exploitation attempts within hours (SecurityWeek). On 14 September a stolen Cloudflare API key let attackers inject ClickFix "paste this command" scripts through the email-marketing company Brevo for about five and a half hours, including onto customers' websites (BleepingComputer). Compromised GitHub Actions came back online on 16 September with their malicious tags still in place.

10. Breaches reached the income statement. Boston Scientific told the SEC on 8 September that an intrusion which began on 25 August was "likely to have a material impact" on its 2026 results, after it disrupted manufacturing and the shipping of orders (SEC). IDScan confirmed a breach on 10 September, and more than 153 million scanned driving licences were found for sale (TechCrunch). Scans are not people, and the number of individuals affected is unclear. The crypto exchange Bitget lost $351.6 million by its own figure, or $387.5 million by its chief executive's first estimate, on 24 September (BleepingComputer).

Cloud, compute and chips

11. Oracle's backlog reached $664 billion, and its cash flow went negative (10 September). Revenue rose 30% to $19.3 billion and its infrastructure cloud grew 121%. It delivered 850 megawatts of capacity and more than 300,000 GPUs in a quarter. Capital spending of $28.5 billion left free cash flow at about minus $5.4 billion (Oracle 8-K). As a new round of layoffs began, its 10-Q raised the restructuring estimate to about $2.8 billion, and an analyst noted that it names AI integration as a driver for the first time (CIO.com).

12. GPU capacity became a seller's market (16–24 September). CoreWeave said it is signing three-to-six-month contracts at about $40 million per megawatt of annualised revenue (BusinessWire). The neocloud Nscale's listing filing showed $103.4 billion of contracted value, about 85% of it from Microsoft and Anthropic, against first-half revenue of $140.6 million (SEC; Quartz). Anthropic signed an $11.6 billion, seven-year deal with Akamai (Bloomberg). Backlogs are not new money, though: new deals with stated values in the first half of the month came to about $65 billion. See AI compute got dearer.

13. Data centres were handed the grid bill (16–22 September). The US House passed the Ratepayer Protection Act by 417 votes to 3. It would push states to make loads above 100 megawatts pay the "full, incremental" cost of grid upgrades, though Senate passage before the midterms is seen as unlikely (Utility Dive). North Carolina rejected a $584 million gas turbine justified by data-centre growth, and California signed a seven-bill package on grid costs and energy reporting. FERC refused to let a utility cancel a 1.8-gigawatt data-centre deal backed by a $1 letter of credit, which one commissioner called "an embarrassing legal fiction" (Utility Dive). Texas went the other way and dropped a proposed interconnection fee.

14. Custom chips boomed, and there was no US–China chip thaw. Broadcom's AI revenue rose 221% to $16.7 billion on 2 September. On its earnings call it outlined about $115 billion of AI revenue in fiscal 2027 (Broadcom 8-K). TSMC's August revenue rose 53% year on year, and Huawei pulled its next Ascend AI chip forward to early 2027. The Trump–Xi state visit of 24–26 September reportedly extended the trade truce to 10 January 2027 with no easing of chip controls. That account rests on a bank note and secondary reports, not an official readout (Saxo).

Policy

15. The EU's 24-hour vulnerability clock started (11–12 September). Cyber Resilience Act reporting now requires manufacturers of hardware and software sold in the EU to give an early warning within 24 hours of an actively exploited vulnerability or severe incident, including for products already on the market. From the next day, the Data Act requires new connected products to make their data accessible by design (European Commission; Faegre Drinker). See the EU's 24-hour clock.

16. Google had a month of rulings. A US judge refused to break up Google's ad-tech business on 2 September. The opinion unsealed on 16 September instead requires interoperability with the open-source Prebid framework, bid-data sharing with publishers and a six-year monitor (US DOJ). Ireland's Data Protection Commission fined Google €403 million over location data on 21 September (DPC). On 23 September the UK's CMA proposed adding AI assistants such as ChatGPT and Perplexity to Google's search choice screens, with consultation until 9 October (CMA).

17. Washington and Sacramento moved in opposite directions. The White House issued no AI executive order in the month. It extended the $100,000 H-1B petition requirement to September 2027 on 18 September (White House). On 24 September the Justice Department asked to join X's court challenge to the EU's first Digital Services Act fine (US DOJ). The same day, the FTC voted to consider making platforms answer for impersonation-scam ads (FTC). California signed laws on independent AI verification and auditors and on AI-generated performers in adverts. On 18 September Governor Newsom ordered work on a "kill switch" for frontier models (Governor of California).

Work

18. Seat licences gave way to meters (16–25 September). Microsoft's "new Copilot" keeps chat in the per-user licence and moves its agents to usage-based billing, with no published rate card. From 2 November, new Copilot Business licences bought through resellers have pay-as-you-go billing on by default (Microsoft Partner Center). At Dreamforce, Salesforce launched agents and an "AIforce" layer without prices. Its platform was then degraded worldwide for about eleven hours on 16 September, the event's second day (The Register). See Copilot's move to the meter.

Consumer technology

19. iPhones got dearer and Siri got an AI brain with conditions (9–18 September). The iPhone 18 Pro starts at $1,199 and the Pro Max at $1,299, both on sale from 18 September. Apple's first foldable, the iPhone Duo, starts at $1,999, with pre-orders from 16 October. Apple also raised the price of the older iPhones it still sells by $100 each, blaming memory-chip costs (Apple; MacRumors). iOS 27 arrived on 14 September with an AI Siri. It is a beta, English-only, not on iPhones in the EU or mainland China, and due to get usage limits and paid extra access (MacRumors). See the new Siri, explained.

20. Smart glasses split and robotaxis scaled (1–24 September). At Meta Connect on 23 September, the Ray-Ban Meta Gen 3 went on sale at $449. Meta also announced a camera-free Ray-Ban Meta Audio at $349, shipping 13 October, which Engadget reads as an answer to privacy criticism (Meta; Engadget). Waymo opened four more US cities, reaching 15, and was averaging about 500,000 paid rides a week by 24 September (TechCrunch). Safety scrutiny grew alongside it: US regulators opened reviews of Tesla's Cybercab certification and of about 30,000 comma.ai driver-assist devices.

What to take from it

Falling unit prices and rising bills now coexist. Budgets built on per-token list prices will miss the meters, the expiring introductory prices, and the grid and GPU costs being passed down the chain.

The evidence for AI's agency ran behind the claims. The best-documented damage came from ordinary failures: appliances exploited before or just after patches, stolen API keys, spoofed phone numbers and a single login service. The AI-agent stories were real, but most rest on one company's telemetry.

Count carefully. Two distortions ran through the month's coverage. One inflated AI's autonomy. The other added up multi-year backlogs as if they were new money. Naming the operator and the counting method would have caught both.

Sources and scope

This roundup covers 1 to 26 September 2026 and is ByteNib editorial analysis. Facts are attributed to their sources inline: company filings and releases, government advisories and court documents, and reporting from outlets including Bloomberg, TechCrunch, BleepingComputer, The Register, Utility Dive and MacRumors. Items resting on anonymous sources, a single outlet or a company's own claims are labelled as such. Some primary sources, including OpenAI's website and several paywalled reports, could not be read directly, and figures from them are attributed as reported. ByteNib uses AI tools, including Anthropic's Claude, in research and drafting under its editorial standards. Anthropic appears in several items above, and its claims are attributed to it.

Continue exploring: IT Leadership analysis, the related implementation tutorial, and the structured learning path.