AI and IT News Recap: July 23, 2026: AMD Bets Up to $5B on Anthropic, OpenAI Ships 'Presence,' and a Critical SharePoint Bug Hits CISA's Must-Patch List
By Noah Smith, Owner & Consultant, KeyChange Technologies · July 23, 2026
Your fast, candid read on the AI and IT news that matters for business owners. Today: AMD writes a nine-figure check to Anthropic, OpenAI turns its models into a managed agent product, and defenders get a fresh batch of critical bugs to patch before the weekend.
📌 The AI and IT news at a glance
- 🔝 AMD commits up to $5B to Anthropic and 2 gigawatts of GPUs to keep Claude fed with compute.
- OpenAI ships "Presence," a managed platform for production voice and chat agents.
- A critical SharePoint RCE (CVE-2026-50522) lands on CISA's must-patch list after a public exploit appeared.
- An Adobe Acrobat browser extension flaw exposed WhatsApp Web chats across ~329 million installs.
- Hackers sat inside South Korea's diplomat training system for ~10 months, exposing roughly 10,000 diplomats.
- Swiss rail maker Stadler refused a 2.3M ransom and called the police instead.
- Anthropic wired its Economic Index into Claude so you can just ask about AI's impact on jobs.
🤖 AI
AMD bets up to $5 billion on Anthropic and a 2-gigawatt GPU buildout
AMD and Anthropic announced a strategic partnership on July 22 to deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in AMD's "Helios" rack-scale systems, with the first gigawatt coming online in the first half of 2027. The deal pairs the hardware with an engineering collaboration: Anthropic will use Claude to help optimize workloads and speed up AMD's ROCm software stack, AMD will roll Claude out across its own engineering and product teams, and AMD has committed to a strategic equity investment of up to $5 billion in Anthropic.
The subtext is the story of the year in AI: compute is the constraint, and everyone is locking up capacity years in advance. Anthropic's chief compute officer Tom Brown framed it plainly, saying the goal is to secure the capacity Claude needs and run "the right workloads to the right hardware" across a diversified fleet. For AMD, landing a frontier lab at gigawatt scale is a direct shot at NVIDIA's near-monopoly on AI training silicon, and it comes just two days after AMD announced a similar Helios expansion with Microsoft's Azure.
In short: AMD and Anthropic signed a partnership to deploy up to 2 gigawatts of MI450 GPUs, with AMD investing up to $5 billion in Anthropic.
What it means for your business: The AI you rent is increasingly gated by who controls the chips, so expect the model vendors with locked-in compute to be the ones who keep shipping, keep prices moving, and stay available when demand spikes.
My take: Watch the equity piece more than the gigawatts. When your GPU supplier also owns a chunk of your biggest customer, "diversifying away from NVIDIA" starts to look like building a second dependency rather than removing one. Good for resilience, but it is not a free lunch, and these mega-deals are why your model bill is unlikely to fall as fast as the marketing suggests.
OpenAI ships "Presence," a managed platform for production AI agents
On July 22, OpenAI introduced Presence, a deployed enterprise product for building and governing AI agents across voice and chat. Rather than handing companies raw model access and leaving the guardrails to them, Presence bundles company-defined policies, escalation rules, simulations, evaluations, and a set of approved actions, then uses OpenAI's Codex coding agent to review real interactions and suggest improvements that staff approve before they go live. OpenAI says it already runs Presence on its own English-language phone support line, where it resolves about 75% of inbound calls without a human.
The pitch is aimed squarely at the gap between a flashy agent demo and something you would trust to talk to customers unsupervised. Early testers include Spanish banking group BBVA, which is exploring AI voice support in Mexico, and Japan's SoftBank, testing natural Japanese conversations. It is a notable strategic shift: OpenAI is moving up the stack from selling tokens to selling a managed system, which puts it in more direct competition with the customer-experience platforms and integrators its own API used to power.
In short: OpenAI launched Presence, a managed platform for deploying guardrailed voice and chat agents, already handling 75% of its own support calls.
What it means for your business: If you have been waiting to put an AI agent on your phone lines or support chat, the "safe enough to deploy" tooling is arriving, but so is deeper lock-in to whichever vendor's stack you build on.
My take: The 75% self-reported deflection number is the part to pressure-test, because "resolved without a human" and "resolved well" are not the same metric. Still, the honest read is that the hard part of agents was never the model, it was the policies, evals, and rollback story around it. That is exactly what Presence is selling, which tells you where the real work is.
Source: https://openai.com/index/introducing-openai-presence/
🛡️ IT and security
A critical SharePoint bug (CVE-2026-50522) hits CISA's must-patch list
CISA added CVE-2026-50522 to its Known Exploited Vulnerabilities catalog on July 22, giving federal agencies until July 25 to fix it. The flaw is a critical (CVSS 9.8) unsafe-deserialization bug in on-premises SharePoint that lets an unauthenticated attacker run code over the network, and it affects SharePoint Enterprise Server 2016, Server 2019, and Subscription Edition. Active exploitation kicked off shortly after a public proof-of-concept appeared, turning a patch-when-you-can item into a patch-now emergency.
This is the latest entry in a brutal month for SharePoint admins, following two other Server flaws from the July Patch Tuesday cycle that already saw attacks. On-prem SharePoint remains a favorite target because it often sits deep inside corporate networks with wide access, so a single unauthenticated RCE can become a foothold for lateral movement and domain compromise.
In short: CISA flagged an actively exploited critical SharePoint RCE (CVE-2026-50522) and set a July 25 federal patch deadline.
What it means for your business: If you run SharePoint on your own servers, this is a drop-everything patch, and if a vendor or partner runs it on your behalf, ask them today whether they have applied it.
My take: The pattern here matters more than the single CVE. Three exploited SharePoint bugs in a month is not bad luck, it is a signal that on-prem SharePoint has become a soft target. If you are still self-hosting it, this month is your cue to plan the migration or the extra network isolation, not just to apply another patch.
Source: https://thehackernews.com/2026/07/critical-sharepoint-rce-cve-2026-50522.html
An Adobe Acrobat browser extension flaw exposed WhatsApp Web chats
Researchers at Guardio Labs disclosed a flaw, dubbed HermeticReader and tracked as CVE-2026-48294, in the Adobe Acrobat extension for Chrome, which is installed in roughly 329 million browsers. The bug is a cross-origin data-disclosure issue: by abusing the extension's internal messaging engine and predictable tab IDs, a malicious website could quietly send commands to a victim's open WhatsApp Web tab and read chat lists, contact names, and message contents. The attack needed only a single visit to a booby-trapped page and no session cookies or logins.
Adobe has fixed the issue in extension version 26.5.2.3, which is being pushed to users automatically, so most people are patched without doing anything. The episode is a useful reminder that browser extensions run with a lot of privilege and quietly bridge the sites you have open, which makes a popular, trusted-brand extension an attractive single point of failure.
In short: A now-patched Adobe Acrobat Chrome extension flaw (CVE-2026-48294) let malicious sites read users' WhatsApp Web conversations.
What it means for your business: Extensions are an underrated slice of your attack surface, and a widely installed one can expose staff communications on completely unrelated sites.
My take: The fix is automatic, so the action item is not "patch this one bug," it is "know what extensions your team runs." Most companies have zero inventory of browser add-ons, yet each one can read and act on the pages your people open. A quick audit and an allowlist buys you more safety than reacting to each headline.
Hackers sat inside South Korea's diplomat training system for months
South Korea's Foreign Ministry disclosed that an online education system run by the Korea National Diplomatic Academy was accessed by an unidentified attacker for roughly 10 months, with a database of about 10,000 current and retired diplomats believed to have been exposed. Officials said the intruder reached the server around April or May 2025 and returned repeatedly until early this year, and intelligence authorities are investigating a possible North Korea link. The ministry also acknowledged it delayed public disclosure while it worked to contain the security risk.
The scale and dwell time are the alarming part. Nearly a year of undetected access to a system holding personal data on essentially the entire diplomatic corps is the kind of breach that fuels years of follow-on targeting and social engineering. It also underscores how "peripheral" systems like a training or e-learning portal often hold sensitive data while getting a fraction of the monitoring that core systems receive.
In short: An unknown attacker had ~10 months of access to a South Korean diplomatic academy system, exposing data on about 10,000 diplomats.
What it means for your business: Your secondary systems, such as training portals, HR tools, and old microsites, are exactly where long, quiet breaches happen, because nobody is watching them.
My take: Ten months undetected is the headline, and dwell time is the metric every owner should care about. It is not whether you will be probed, it is how fast you notice. If you cannot say how you would detect an intruder living in a forgotten portal for a month, that gap is worth more attention than the next shiny tool.
Source: https://therecord.media/south-korea-cyberattack-foreign-ministry
Swiss rail maker Stadler refuses a
2.3M ransomSwiss train manufacturer Stadler Rail said the Everest extortion gang breached a data-exchange platform it shares with a supplier, stole technical files, and demanded about
2.3 million (roughly CHF 10 million). Stadler publicly refused to pay, said it will not negotiate under any circumstances, and filed a criminal complaint with the Thurgau cantonal police. The company says the stolen material was non-safety-relevant technical data belonging to the supplier, with no impact on personal data, its own IT systems, production lines, or trains in service worldwide.Everest is a data-theft extortion crew that dropped file-encrypting ransomware years ago in favor of pure "pay us or we leak it" pressure. Notably, the group has not yet publicly claimed the attack or listed Stadler on its leak site, so the standoff is still playing out. The breach also lands through a supplier's shared platform, a reminder that third-party connections are a common way attackers reach an otherwise well-defended company.
In short: Stadler Rail refused a
2.3 million Everest ransom after a supplier-shared platform was breached, and reported it to police.What it means for your business: Refusing to pay is more viable when you can prove the stolen data is limited, which only happens if you know what a partner platform actually holds before something goes wrong.
My take: Stadler's confidence came from scope: they could say precisely what was and was not taken, which is what let them say no. Most small and mid-size firms cannot make that statement, and that uncertainty is exactly what ransom crews sell back to you. The lesson is boring but real, which is that mapping what data lives in each vendor connection is what buys you the option to refuse.
🧰 New tooling for business and vibe coders
Anthropic wires its Economic Index into Claude
Also on July 22, Anthropic launched an Anthropic Economic Index connector for Claude, letting you ask questions about how AI is showing up across occupations, tasks, and trends and get answers grounded in the Index's underlying data rather than the model's general knowledge. The Economic Index is Anthropic's ongoing research effort to measure how people actually use AI at work, and the connector turns that dataset into something you can query conversationally, with the source data available to dig into.
It is a small release next to a $5 billion partnership, but it points at a useful pattern for business users: connectors that ground a chatbot in a specific, trustworthy dataset instead of letting it freewheel. If you are trying to reason about where AI fits in your own staffing or workflows, having the numbers one question away beats hunting through a PDF report.
In short: Anthropic released a connector that lets Claude answer questions about AI's economic and workforce impact using its Economic Index data.
What it means for your business: It is a free, low-effort way to sanity-check claims about which jobs and tasks AI is really touching before you make a staffing or automation decision.
My take: The interesting bit is not the dataset, it is the shape. Grounding an assistant in a named, auditable source is how you get useful answers instead of confident guesses, and it is the same trick worth applying to your own internal data. Treat this one as a template, not just a research toy.
Source: https://www.anthropic.com/news/anthropic-economic-index-connector
Miss yesterday's rundown? Catch up on the July 22, 2026 AI and IT News Recap.