AI and IT News Recap: July 10, 2026: OpenAI's ChatGPT Work Ships Finished Documents, Grok 4.5 Undercuts the Field, and a Microsoft Defender Zero-Day Gets Patched

By Noah Smith, Owner & Consultant, KeyChange Technologies · July 10, 2026

Your no-spin AI and IT news roundup for July 10, 2026. The theme today is agents that finish the job: OpenAI shipped ChatGPT Work, a "super app" agent that goes off and produces the actual documents, slides, and websites, while xAI dropped Grok 4.5 at a price built to undercut everyone. On the security side, Microsoft patched a nasty Defender zero-day, two big data exposures came to light (a hacker claims 35GB of Accenture source code, and an auto insurer spilled nearly 7 million driver's licenses), and INTERPOL hauled in almost 6,000 scam suspects. Here is what actually matters for your business.

📌 The AI and IT news at a glance

• 🚀 OpenAI launches ChatGPT Work, an autonomous agent that spends hours producing finished docs, slides, sheets, and websites across your apps.
• ⚡ xAI ships Grok 4.5, an "Opus-class" model trained on real coding-tool data and priced well below Claude Opus 4.8 and GPT-5.6.
• 🛡️ Microsoft patches RoguePlanet (CVE-2026-50656), a Defender zero-day that handed attackers SYSTEM access on fully updated Windows.
• 💼 Accenture confirms a breach after a hacker offered 35GB of stolen source code, RSA/SSH keys, and Azure access tokens for sale.
• 🪪 AssuranceAmerica breach exposes ~6.9 million drivers, the largest known spill of US driver's license numbers so far this year.
• 🌐 INTERPOL's Operation First Light 2026 ends with 5,811 arrests and 93 million frozen across 97 countries.

🔝 Top story: OpenAI launches ChatGPT Work, an agent that ships finished work

On Thursday, July 9, OpenAI unveiled ChatGPT Work, its long-teased "super app" that fuses the ChatGPT chatbot with its Codex coding engine and points the whole thing at white-collar busywork. Instead of just answering questions, ChatGPT Work is built to take action across your apps and files and spend hours, if needed, producing finished materials: documents, spreadsheets, slide decks, and even working web apps. It runs on GPT-5.6, supports task scheduling so you can kick off a job from your phone and check progress later from a browser or the new desktop app, and includes built-in computer-use so it can click around your apps and the web to get things done. A hosted-websites feature lets you build and publish sites directly from the tool.

The rollout started July 9 on web and mobile for Pro, Enterprise, and Edu users, with Plus and Business users following over the next few days. The framing is unmistakable: this is OpenAI's direct answer to Anthropic's Claude Cowork, and it lands in the middle of an intensifying race to sell autonomous agents that complete real work with minimal hand-holding rather than chatbots that just draft and suggest.

In short: OpenAI released ChatGPT Work, a GPT-5.6-powered agent that autonomously produces finished documents, spreadsheets, slides, and websites across your apps, starting July 9 for Pro, Enterprise, and Edu.

What it means for your business: This is the clearest sign yet that the useful unit of AI is shifting from "answer" to "deliverable," so the highest-value place to point it is the multi-step admin work (reports, reconciliations, first-draft decks) that eats your team's hours, not one-off questions.

My take: The pitch is seductive and the caution is the same as ever. An agent that runs for hours and touches your files and apps is genuinely useful and genuinely riskier than a chatbot, because now a confident mistake ships as a finished document or a live website rather than a suggestion you review. Start it on low-stakes, reversible work, keep a human checkpoint before anything goes to a client or gets published, and watch what permissions you hand it. The productivity is real; so is the blast radius when it gets something wrong.

Source: OpenAI unveils long-awaited "super app" as rivalry with Anthropic intensifies (Reuters via U.S. News)

🤖 AI

xAI ships Grok 4.5 at a price built to undercut the field

xAI released Grok 4.5 this week, with Elon Musk opening it to the public on July 9 after a short beta. The company positions it as its model for coding, agentic tasks, and knowledge work, built on a 1.5-trillion-parameter foundation and trained partly on real session data from the Cursor coding tool. Musk called it "an Opus-class model, but faster, more token-efficient and lower cost," and independent trackers put it fourth on the Artificial Analysis intelligence index, above every open-weight model.

The number that will get attention is price: Grok 4.5 runs per million input tokens and $6 per million output tokens, with a 500K-token context window. That is well under Claude Opus 4.8 (roughly $5 and 5) and OpenAI's new GPT-5.6 Sol tier, and it is the second frontier-class launch in two days after GPT-5.6 went public. The pattern is a genuine price war at the top of the market.

In short: xAI's Grok 4.5, publicly available July 9, is an Opus-class model with a 500K context window priced at /$6 per million tokens, undercutting Claude Opus 4.8 and GPT-5.6.

What it means for your business: Frontier capability keeps getting cheaper, so any workflow you priced out six months ago is worth re-costing; the smart move is to treat the model as a swappable component and route each task to the cheapest one that clears your quality bar.

My take: The headline capability claims all come from xAI's own benchmarks, so treat "Opus-class" as a marketing claim until you have run it on your own tasks. What is not in dispute is the pricing, and that is the real story: with GPT-5.6, Opus 4.8, and now Grok 4.5 all fighting on cost, the leverage sits with buyers for the first time in a while. Build so you can switch providers without a rewrite, and you get to keep pocketing these price cuts instead of being locked to one vendor's roadmap.

Source: Introducing Grok 4.5 (xAI)

🛡️ IT and security

Microsoft patches RoguePlanet, a Defender zero-day that handed over SYSTEM

Microsoft has shipped a fix for RoguePlanet (CVE-2026-50656), a zero-day elevation-of-privilege flaw in the Microsoft Malware Protection Engine, the core scanner behind Microsoft Defender. The bug is a race condition that let a local attacker spawn a command prompt with SYSTEM privileges on a fully patched Windows 10 or Windows 11 machine, effectively turning the security tool itself into the way in. Microsoft addressed it by updating the scanning engine (version 1.1.26060.3008), which rolls out automatically to most systems rather than through the usual monthly patch cycle.

The awkward part is the pedigree: this flaw was disclosed shortly after June's Patch Tuesday and, per reporting, was tied to active exploitation before the fix landed. Because the malware engine updates itself in the background, most machines will get protected without admin action, but "most" is not "all," and organizations that manage or delay engine updates should confirm they are on the patched build.

In short: Microsoft patched RoguePlanet (CVE-2026-50656), a Defender zero-day that gave local attackers SYSTEM-level control on fully updated Windows, by pushing a new Malware Protection Engine build.

What it means for your business: The fix ships automatically for most Windows machines, so your main job is verification: confirm your endpoints are running Malware Protection Engine 1.1.26060.3008 or later, and pay special attention to any systems where engine updates are centrally controlled or air-gapped.

My take: There is a grim irony in the tool you install to catch attackers being the thing that hands them the keys, and it is a useful reminder that security software is software, with its own bugs and its own privileged access. The saving grace here is the auto-updating engine, which quietly patches the long tail of machines that never get touched by IT. Do not assume, though: spend ten minutes confirming the build number on your managed fleet, because "it updates itself" is exactly the belief attackers count on.

Source: Microsoft patches RoguePlanet Defender zero-day vulnerability (BleepingComputer)

Accenture confirms a breach after a hacker offers 35GB of source code

Accenture, one of the world's largest IT consultancies, has confirmed a security incident after a threat actor using the handle "888" posted on a cybercrime forum offering "just over 35GB of source code" and other data for sale. The seller claims to have exfiltrated source code, RSA keys, SSH keys, Azure personal access tokens, Azure Storage access keys, and configuration files, and included a screenshot said to show a private Azure DevOps repository on an Accenture production URL. Accenture confirmed it is aware of the matter, described it as "isolated," and said it "remediated its source," but did not confirm the attacker's specific claims about the volume or type of data taken, how access was gained, or whether client data was involved.

The reason to watch this one is the nature of what is allegedly for sale. Source code plus live cloud credentials (access tokens and storage keys) is a supply-chain problem, not just an embarrassment, because those secrets can unlock downstream systems and customer environments if they were valid and not rotated.

In short: Accenture confirmed an "isolated" breach after a hacker advertised 35GB of allegedly stolen source code, SSH/RSA keys, and Azure access tokens, though the firm did not verify the specifics.

What it means for your business: When a vendor's source code and cloud keys leak, the risk flows downstream, so this is a prompt to ask your critical suppliers how they detect and rotate exposed secrets, and to make sure your own repos are not storing live credentials in the first place.

My take: Set aside the disputed 35GB figure; the detail that matters is the claim of leaked access tokens and storage keys. Hardcoded secrets in source control are one of the most common and most preventable breach multipliers, because one stolen repo becomes a skeleton key. Treat this as a nudge to audit your own codebase for embedded credentials, turn on secret scanning, and rotate anything that has been sitting in a repo for years. You do not want to be the next company explaining an "isolated matter."

Source: Accenture confirms breach after hacker offers stolen data for sale (BleepingComputer)

An auto insurer just exposed nearly 7 million driver's licenses

AssuranceAmerica, an Atlanta-based auto insurer, is notifying about 6.9 million people that their personal data was stolen, making it the largest known exposure of US driver's license numbers reported this year. The stolen files include names, contact information, driver's license numbers, and auto policy, vehicle, and claims details. The company says it spotted suspicious activity on March 17, 2026, after an attacker compromised an employee's credentials the day before and copied data from its environment; the forensic review wrapped on June 15, and notifications are going out now.

Driver's license numbers are a slow-burn problem: unlike a card number, you cannot easily reissue them, and they are prime raw material for identity theft and synthetic-identity fraud that can surface months or years later. The four-month gap between discovery and notification is also a reminder that "we contained it" and "we told the people affected" can be a long way apart.

In short: AssuranceAmerica is notifying ~6.9 million people that a March intrusion, traced to stolen employee credentials, exposed their driver's license numbers and insurance details.

What it means for your business: The root cause was one compromised employee login, which is the most common way breaches start, so the takeaways are boring and effective: enforce phishing-resistant MFA, limit how much data any single account can reach, and watch for unusual bulk data access.

My take: Two lessons stack up here. First, a single set of stolen credentials copied millions of records, which tells you the real control that matters is not just "did we have MFA" but "how much can one account touch before something flags it." Second, driver's license data does not expire, so the harm from this outlives the news cycle. If you handle customer PII, assume any one employee account can be phished, and design so that a single compromised login cannot quietly walk out with your whole database.

Source: AssuranceAmerica data breach exposes records of 6.9 million drivers (BleepingComputer)

INTERPOL's global scam bust: 5,811 arrests and 93 million frozen

INTERPOL announced on July 9 that Operation First Light 2026, a four-month push across 97 countries, led to 5,811 arrests and froze 93 million tied to social-engineering fraud. Investigators worked roughly 150,000 cases, froze more than 31,000 bank accounts, and identified over 142,000 victims. The targeted scams are exactly the ones that hit businesses: business email compromise, investment and romance cons, impersonation, and the money-laundering networks that move the proceeds, including one Thai suspect who allegedly laundered more than