AI Deals, Security Flaws, and Microsoft’s New Assistants
Meta chatbot vulnerability, OpenAI on AWS, Buffett’s billions for AI infrastructure, and new assistants from Google and Microsoft: today’s most important AI news.
Inhaltsverzeichnis
Today once again shows very clearly why AI news right now is not just “exciting,” but also quite practical: It’s about real security risks, new enterprise offerings, and the question of who will pay for the infrastructure behind the next AI wave. In short: less show, more consequences.
And yes, if an AI chatbot suddenly changes accounts, that’s not a feature, but a bug with PR risk.
🔐 Meta chatbot as an entry point: Instagram takeovers despite 2FA
Hackers took over prominent Instagram accounts by simply asking Meta’s AI support chatbot to change the email address. It sounds absurd, but that’s exactly the problem: when an automated support channel is trusted too much, convenience quickly turns into a security hole. Particularly serious is that two-factor authentication was bypassed in the process — exactly the protection layer many people rely on.
Why does this matter? Because it shows how vulnerable AI-powered support systems can be when they’re directly tied to identity and account processes. This is not just a social media problem, but a pattern for many platforms: AI can assist, but it must not administer blindly. According to security researchers, another exploit is already said to be circulating on Telegram, even though Meta has patched the vulnerability. Unfortunately, this is the moment when “fixed quickly” and “already circulating too late” can be true at the same time.
Source: The Decoder
☁️ OpenAI models now available directly via AWS
OpenAI is bringing GPT-5.5, GPT-5.4, and Codex to AWS via Amazon Bedrock. For companies, that’s a pretty big step because it lets them use AI models within their existing cloud and compliance environment instead of rebuilding workflows from scratch. According to the report, pricing remains at the same level as using OpenAI directly, and existing AWS contracts are supposed to count toward it. That makes adoption significantly easier for enterprise customers.
Why is this important? Because the market is continuing to shift from “which model is best?” to “where can it be operated best?” For many companies, the bottleneck is not the model itself, but integration, governance, regional availability, and procurement. The fact that the models initially run only in the U.S. slows down international rollouts, but the direction is clear: AI is increasingly becoming a cloud infrastructure question. So if you handle cloud purchasing today, you might want to mentally prepare for a few extra lines in the # budget.
Source: The Decoder
💸 Buffett invests 10 billion in Alphabet’s AI offensive
Warren Buffett is investing $10 billion in Alphabet — at precisely the moment when the company is massively expanding its AI infrastructure. According to the report, Alphabet plans capital expenditures of up to $190 billion for 2026, with even more expected in 2027. That’s not a small bet, but an infrastructure race on Olympic scale.
What does that mean for the market? First: AI is no longer just software, but above all compute, networks, energy, and data centers. Second: when an investor like Buffett joins this offensive, it sends a signal to the entire market — not necessarily that everything is cheap, but that long-term demand for AI infrastructure is being viewed as real. For companies planning their own models or AI products, this also means: capacity is strategic. Anyone aiming to scale today needs not only good prompts, but reliable compute planning. And probably a few good contacts at power utilities as well.
Source: The Decoder
📞 Google Phone warns about AI impersonations
Google is expanding its Phone app with a feature designed to warn you about calls where scammers appear to be imitating the identity of a contact. If an incoming number looks familiar but shows suspicious behavior, the app flags the call as potentially fraudulent. This is part of a broader Android update in June.
The background is clear: AI makes fraud more convincing. Voices can be cloned, identities forged, and social closeness simulated — in a quality that often defeats classic fraud warnings. Such a protection feature is therefore useful because it addresses the exact point where users make an everyday decision: “answer or not?” For smartphone security, this is an important building block, even if it’s certainly not a cure-all. But a small hint at the right moment can be worth more than the hundredth warning email in the spam folder.
Source: The Verge
🧩 OpenAI turns Codex into a general-purpose app for non-developers
OpenAI is continuing to expand Codex — this time with role-specific plugins for areas such as data analysis, sales, or investment banking. What’s exciting is that, according to OpenAI, five million people use the tool weekly, and one-fifth of them are not developers. This user group is even growing three times faster than the developer base. Codex is thus increasingly evolving from a developer tool into a general work app for knowledge workers.
Why does this matter? Because a clear trend is emerging here: AI tools gain value not only through “more model power,” but through context and task orientation. If a tool knows whether you’re in a sales meeting, financial analysis, or a data check, it becomes more useful — and more dangerous if permissions, data access, or hallucinations are not properly constrained. That’s exactly why role-based plugins are so interesting. They make AI more practical, but also a topic for IT, compliance, and #.
Source: The Decoder
🧠 Microsoft Scout: always-on assistant for Microsoft 365
Microsoft is working on Scout, a new AI assistant intended to integrate deeply into Microsoft 365. Unlike Copilot, which is embedded in apps like Outlook or Teams, Scout appears to be designed more as a persistent, cross-cutting assistant — basically a digital colleague that organizes calendars, drafts emails, prepares expenses, and accompanies workflows. For companies, this is interesting because the assistant can help not just in one app, but across the Microsoft stack.
This is a logical next step: instead of point-by-point support, Microsoft seems to want to build a more persistent work engine. That increases usefulness, but also raises the bar for permissions, transparency, and control. Because an always-on assistant that sees a lot can also do a lot wrong if roles and approvals aren’t clearly defined. For enterprise teams, that’s good news — as long as productivity doesn’t get accidentally blocked by the company’s own IT policy.
Source: The Verge
🛠️ Tool tip of the day
If you want to build AI workflows, assistant functions, and automation for your team, it’s worth taking a look at modern workflow tools with clean integration logic. Especially with LLM-powered processes, it’s important to clearly separate actions, approvals, and data access — otherwise “automation” turns into chaos faster than planned. For teams looking to get started cleanly, a structured approach with # is often worth its weight in gold.
Don’t want to miss any news? Subscribe to the newsletter