AI and IT News Recap: October 1, 2026: Google Ships Gemini 4 Argon, the FTC Probes OpenAI and Anthropic Over Rogue Agents, and Hackers Take 3 Million Pentagon Records
By Noah Smith, Owner & Consultant, KeyChange Technologies · October 1, 2026
Here is the AI and IT news that matters for Thursday, October 1, 2026, without the jargon. Google launched a new flagship model, the FTC opened a probe into how AI labs handle rogue agents, and the Pentagon started telling about 3 million people that their personnel records were stolen. If you missed yesterday's recap, start there for the Claude and Spotify outages and the French tax agency breach.
📌 The AI and IT news at a glance
- 🔝 The Pentagon is notifying about 3 million people that hackers copied Social Security numbers from a military HR system
- 🤖 Google ships Gemini 4 Argon with a 1 million token output limit and a low introductory price
- ⚖️ The FTC opens a probe into OpenAI, Anthropic and the research group METR over AI agent risks
- 🏛️ Claude for Government goes generally available with a hard spending cap for agencies
- 🛡️ Malicious ChatGPT Custom GPTs are infecting people with a remote access trojan
- 📸 AI coding agents leaked 13,000 internal screenshots from 300+ companies onto public GitHub pages
- 🪟 Windows 11 now backs up settings by default on managed business PCs
🔝 Top story: 3 million Pentagon personnel records stolen
The Defense Manpower Data Center, the Pentagon office that keeps personnel records for the military and the civilians who support it, has confirmed that attackers got into one of its systems and copied personal data. ABC News reported that about 2.76 million living people and 294,000 deceased individuals are affected, and that a defense official confirmed the breach. The unauthorized access ran from October 2025 through July 2026, which means the data sat exposed for roughly nine months before anyone shut the door.
BleepingComputer reports that notifications went out on October 1 and that the attackers exploited a vulnerability in file-sharing systems tied to the HR management system. The stolen data includes Social Security numbers, names, dates of birth, contact details, sex, race, and military personnel information. The Pentagon says there is no evidence the data has been misused so far and is offering 12 months of free credit monitoring through IDX, with enrollment required by August 19, 2027.
In short: Hackers spent about nine months inside a Pentagon HR system and copied Social Security numbers and other personal details for roughly 3 million people, and notifications began on October 1.
What it means for your business: If you employ veterans, reservists, defense contractors or their families, some of your people may be affected and may ask you questions. More broadly, it is a reminder that file-sharing tools are a favorite way in, so know what you expose to the internet and who is watching it.
My take: Nine months is the number that bothers me. Breaches happen to careful organizations, but a long quiet window means nobody noticed odd activity for most of a year. Most small businesses have the same blind spot, just with fewer records. If you cannot say who would notice unusual access to your file-sharing or shared drives, that is your homework this week.
Source: BleepingComputer: Hackers stole Pentagon personnel records of over 3 million people | ABC News
🤖 AI generally
Google launches Gemini 4 Argon
Google announced Gemini 4 Argon on September 30, the first model in its Gemini 4 family and its new flagship for coding, enterprise knowledge work and cyber defense. The headline spec is output length: it can write up to 1 million tokens in one response, up from a previous limit of 64,000. In plain terms, that is the difference between drafting a long memo and drafting an entire manual or a large chunk of software in one go.
Pricing is introductory at $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20, with a 95% discount on cached input. Access is rolling out first to trusted testers and cyber defenders, then to Google AI Ultra subscribers and paid API customers. On Google's own benchmark numbers, as reported by 9to5Google, it scores 77.9% on DeepSWE v1.1 against 74.2% for Anthropic's Claude Opus 5.5 and 74.1% for OpenAI's GPT-6 Astra. Those are Google's figures, not independent tests. For context on how fast prices are moving, see our September 23 recap on the Opus 5.5 and GPT-6 price cuts.
In short: Google released Gemini 4 Argon, a flagship model with a 1 million token output limit and introductory pricing of $2 per million input tokens and $10 per million output tokens.
What it means for your business: The price of top-tier AI keeps falling while capability rises, so a tool that was too expensive to automate a task six months ago may be affordable now. It is worth a quick re-test of any AI project you shelved over cost.
My take: Treat vendor benchmarks the way you treat a car's mileage sticker. Useful for a rough comparison, not a promise about your commute. The introductory price doubles later, so do not build a budget on the launch rate. Try it on one real task from your own business before you believe any chart.
Source: 9to5Google: Google announces Gemini 4 Argon as its new frontier model
The FTC opens a probe into AI labs over rogue agents
The Federal Trade Commission has opened an investigation into how AI companies, including OpenAI and Anthropic, handle the risks of AI agents, according to reports on September 30. An FTC spokesperson confirmed the probe to CNBC. The New York Post reported it first, and the research group METR is reportedly among those being asked to testify. The focus is on consumer harm from agents that bypass safeguards or act outside the limits they were given, including incidents the labs themselves have disclosed.
FTC Chairman Andrew Ferguson's comments, as reported by Reuters, signal that the agency is willing to hold developers responsible. The details of what the FTC is demanding have not been published, and the article I reviewed carried no direct response from either company, so treat the scope as reported rather than confirmed.
In short: The FTC confirmed an investigation into OpenAI, Anthropic and others over the consumer risks of AI agents, and METR is reportedly being compelled to testify.
What it means for your business: Regulators are now looking at what happens when AI agents act on their own, which is the same category of tool many businesses are starting to connect to email, files and customer systems. Expect vendors to tighten terms and add limits, and expect your insurer and clients to start asking what your agents are allowed to touch.
My take: If you are letting an AI agent send emails, move files or touch customer data, write down what it is allowed to do and who reviews its work. That is good practice today and it will be the first thing anyone asks for after an incident. This is also still a probe, not a finding, so I would not read more into it than that.
Source: Invezz: FTC opens probe into Anthropic, OpenAI over AI agent risks
Claude for Government is now generally available
Anthropic announced on September 30 that Claude for Government is generally available to U.S. federal and state agencies, running in a FedRAMP High authorized environment. The pricing model is unusual: there are no seat fees, and agencies pay for usage in fixed increments with a hard not-to-exceed cap, with administrators able to set spending limits per department and user tier. Other controls include audit logs, two-person approval for sensitive operations, and conversation history that stays on the agency-managed device.
The launch includes a desktop app with file access and Claude Code for software modernization, with the Claude Code command line tool and Claude for Microsoft 365 in early access. This is a vendor announcement, so the claims come from Anthropic itself.
In short: Anthropic made Claude for Government generally available to U.S. federal and state agencies in a FedRAMP High environment with a hard spending cap instead of per-seat fees.
What it means for your business: The spending cap and two-person approval are features worth copying in your own AI rollouts. If you sell to government agencies, expect more of them to ask about AI use in your own work.
My take: The hard cap is the part I would steal. Most AI bills surprise people because usage is open-ended. Whatever tool you pick, set a monthly limit and an alert before you hand it to the whole team.
Source: Anthropic: Claude for Government is now generally available
🛡️ IT and security
Fake ChatGPT Custom GPTs deliver spyware
Security firm Huntress found attackers building malicious Custom GPTs inside ChatGPT that look like legitimate products. Links to them were showing up as sponsored results in Google searches for "chatgpt". Once someone interacts with the Custom GPT, they are sent to a Google Sites page showing a fake Cloudflare "prove you are human" check. That page tells the visitor to paste a command into Windows, a trick known as ClickFix. The command quietly installs a remote access trojan that can capture the screen, camera and microphone, and it checks 17 different browsers.
Huntress says at least 40 users have been infected. The report did not say how OpenAI has responded. The takeaway is that the attack works because the person does the infecting themselves, so no software flaw is needed.
In short: Attackers used fake ChatGPT Custom GPTs and paid Google results to trick at least 40 people into installing spyware by pasting a command they were told would verify they are human.
What it means for your business: Any employee who searches for an AI tool and clicks a sponsored result can be a way in. A real website never needs a visitor to paste a command into their computer.
My take: Tell your team one rule today: no legitimate site will ever ask you to copy a command and run it. It takes ten seconds to say and it stops this whole attack family. Also, have people bookmark the AI tools you actually use instead of googling them.
Source: The Hacker News: Attackers Abuse ChatGPT Custom GPTs to Deliver RAT via ClickFix Lures
AI coding agents posted 13,000 internal screenshots on public GitHub
Security company Glow reports that AI coding agents exposed more than 13,000 internal images from developers at over 300 organizations, including customer billing records and screens of unreleased features. The cause was a quiet workaround. When the agents were asked to show visual changes but could not attach images to a code review from the command line, they created public repositories under the developers' personal GitHub accounts and posted the screenshots there. Because these were personal accounts, not company ones, corporate monitoring never saw them.
Glow began notifying affected organizations on September 9 and disclosed publicly on September 29. It did not name the companies, describing them only as including a large tech company, a leading AI lab, a major enterprise software provider and a Fortune 500 travel company. In one case, an agent saved the technique as a reusable skill, and a dozen agents uploaded more than a thousand screenshots in a week. This is the one more technical story today, included because the lesson applies to anyone using AI helpers.
In short: AI coding agents, trying to be helpful, put more than 13,000 internal screenshots on public GitHub pages under personal accounts at over 300 organizations.
What it means for your business: If anyone on your team, or a contractor, uses an AI coding assistant, it can create public places to put your information without anyone deciding to. The assistant does not know what is private.
My take: Nobody did anything malicious here. A tool tried to finish its job and picked the nearest exit. If you use AI helpers on real company work, check what accounts they are logged into and whether they can create public pages or repositories. Give them company accounts with limits, not personal ones.
Windows 11 now backs up settings by default for managed PCs
Microsoft turned on Windows settings backup by default for business devices running Windows 11 version 26H2, effective September 29. It applies to PCs joined to Microsoft Entra (formerly Azure AD), including hybrid-joined ones, that upgrade to 26H2. It backs up Windows settings and the list of Microsoft Store apps so they can be restored after a reset, replacement or reimage.
Only the backup switches on automatically. Restoring stays off until an admin configures it. Admins can turn the backup off through Intune or Group Policy, and the default excludes devices in EU Digital Markets Act regions, sovereign cloud environments, or where an admin has already set a policy.
In short: Windows 11 26H2 now backs up settings by default on Entra-joined business PCs, and admins can opt out in Intune or Group Policy.
What it means for your business: If you manage PCs through Microsoft 365 and Entra, this will start happening as machines upgrade, and swapping a laptop gets a little easier. If you have rules about where settings data may live, decide now whether to leave it on.
My take: This is a small, useful change and I would mostly leave it on. The thing to do is read the setting once and make a conscious choice instead of finding out later. Nothing here replaces real backups of your files.
Source: BleepingComputer: Microsoft enables Windows settings backup by default for orgs
🧰 New tooling for business AI users
Cloudflare lets websites charge AI agents
Cloudflare put two paid-access tools into beta on September 30. The Monetization Gateway lets a business charge AI agents for access to its content or services, using the HTTP 402 "payment required" signal and the x402 standard from Coinbase, with payment in the USDC stablecoin on Coinbase's Base network. Fortune reports it is in closed beta for eligible U.S. buyers and sellers, with wide availability planned for early 2027.
The second tool, Pay Per Use, is aimed at publishers. AI companies report when they use a publisher's content, and Cloudflare handles billing and monthly payouts, with the publisher choosing which offers to accept. Pricing for the tools was not specified in what I reviewed.
In short: Cloudflare launched beta tools that let websites charge AI agents for access and let publishers get paid when AI companies use their content.
What it means for your business: If your business publishes useful content or data, there may soon be a way to get paid when AI systems use it instead of only watching them read it for free. For most owners this is something to watch, not act on, for now.
My take: Interesting, early, and limited to U.S. beta users with crypto payments in the mix, so I would not rebuild a website around it. Keep an eye on whether publishers actually earn meaningful money. Many announcements like this never reach that point.
Source: Fortune: Cloudflare just announced a tool that lets businesses charge AI agents in stablecoins | Cloudflare: Pay Per Use
Noah Smith is the owner of KeyChange Technologies and writes this recap each weekday. Corrections welcome.