AI Mornings with Andreas Vig
Your daily AI news briefing in under 10 minutes. New models, product launches, research breakthroughs, and industry shifts, explained clearly, no hype.
AI Mornings with Andreas Vig
OpenAI's Cyber Counterattack & AI Loses Attorney-Client Privilege
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Hey, welcome to AI Mornings with Andreas Vig. It's April 16th, 2026. OpenAI just fired a direct shot at Anthropic with the launch of GPT-5-4, Cyber, a cybersecurity-focused model that takes the exact opposite approach to Anthropic's restrictive Mythos rollout. Remember, Mythos was limited to just 4-0, something whitelisted organizations, mostly tech giants. GPT-5-4, Cyber is available to anyone who passes ID verification through OpenAI's trusted access for cyber initiative. The model can reverse engineer compiled software to identify malware and security flaws. OpenAI researcher Fuad Matin kept it blunt. No one should be in the business of picking winners and losers on who gets to defend their systems. This comes just a week after Treasury Secretary Bessant summoned Wall Street leaders to an emergency briefing on Mythos's hacking capabilities. The cybersecurity model wars are officially on. Now, here's something that affects anyone using AI in legal contexts. A federal judge in New York just issued the first ever ruling on whether AI conversations are protected by attorney-client privilege. The answer is no. The case involved a defendant named Bradley Hepner who was under investigation for securities fraud. After receiving a grand jury subpoena, he used Claude to prepare defense strategy documents. The government wanted access to those AI chats, and Judge Jed Rakoff agreed. His reasoning, Claude isn't an attorney. The communications were shared with a third-party anthropic, and Claude explicitly disclaims providing legal advice. The ruling means if you're discussing legal strategy with an AI, that conversation is not privileged. This is a landmark decision that anyone using AI tools for sensitive matters needs to understand. Nvidia released something interesting today called ICING, the first family of open source AI models designed specifically for quantum computing. These models tackle two of the biggest bottlenecks keeping quantum computers from scaling: calibration and error correction. The calibration model turns what used to be a days-long manual job into something that takes hours. The error decoding model runs at two and a half times the speed with three times the accuracy of current alternatives. Jensen Huang is calling it the operating system of quantum machines. Over 20 institutions are already using it, including Harvard, Fermi Lab, and IonQ. NVIDIA is basically running the same playbook they used for self-driving cars and robotics. Release the OpenAI layer, lock in the ecosystem, own the infrastructure. Alright, a few more things worth knowing about today. OpenAI updated its agents SDK with sandboxing capabilities that let agents run in controlled environments without putting your whole system at risk. They're also adding what they call an in-distribution harness for frontier models, basically a framework for deploying and testing advanced agents working with files and tools in a workspace. Python support is live, TypeScript coming later. It's clearly aimed at enterprises building what they call long horizon tasks, the multi-step complex work that agents are supposed to handle. Adobe officially launched Firefly AI Assistant, which was previewed last year as Project Moonlight. It's an agent that can work across Creative Cloud apps. You describe what you want, and it orchestrates the work in Photoshop, Premiere, Lightroom, and other tools. They're introducing contextual controls, so if you're editing a product photo in a forest, you might get a slider to adjust the foliage. They're also launching skills that combine multiple steps, like automatically cropping and optimizing images for different social platforms. Public beta starts in a few weeks. Anthropic redesigned Claude Codes desktop app around the reality that developers now run multiple AI coding sessions at once. There's a new sidebar for managing sessions, drag and drop panes, and an integrated editor and terminal. They also introduce something called routines that can run AI tasks on schedules, via API, or triggered by GitHub events. The message is clear. Claude Code is becoming a command center for teams of humans and AI agents working together. Two startup stories worth noting. Guitar emerged from stealth with 9 million US dollars led by Venrock. It's focused on AI agents that validate code, targeting the code overload problem where AI-generated code introduces bugs that senior engineers have to fix. And India's Emergent, which raised$70 million in January, launched Wingman, an AI agent that operates through WhatsApp and Telegram. You chat with it to assign tasks, and it works in the background across your email, calendar, and workplace tools. That's it for today. See you tomorrow.