Sec & AI News — 8 May 2026
🟣 Anthropic Signs A Compute Deal With Elon
Anthropic just got higher Claude Code and API limits. The trick: a compute partnership with SpaceX, plus a $200 billion commitment to Google Cloud TPUs. Two months ago Elon was calling Anthropic "missanthropic." Now he's selling Dario the unused 89% of xAI's GPU farm. Enemy of my enemy, billed by the hour. The OpenAI lawsuit is going great.
🟢 Mira Murati Says Sam Lied. Texts Confirm The Vibes.
Former CTO Mira Murati testified under oath that Altman falsely told her legal had cleared a model from the safety board. She checked. They hadn't. Court also surfaced the firing-weekend texts: Sam asks "Can I come in?" Mira: "They don't want you to." And the shocker — Greg Brockman's stake is now worth $30 billion, which is awkward for a nonprofit.
🔵 OpenAI Ships GPT-5.5 Instant. Most Users Won't Notice.
New default model in ChatGPT and Microsoft 365 Copilot. Claims: smarter, more concise, more personalised. Reality: shorter scrollbar, mildly better math, occasionally references your boba history. Not a state-of-the-art jump — an incremental polish on the existing instant model. Available on every paid tier and the free tier. Default upgraded. Move on.
🟡 OpenAI's Voice Models Now Translate Mid-Sentence
Three new realtime models: GPT Realtime 2 (GPT-5-class voice), Realtime Translate (70 input languages → 13 output, live, no waiting for the speaker to finish), Realtime Whisper (streaming speech-to-text). The translation demo is genuinely impressive — listens for the verb, starts translating before the sentence ends. API only for now. ChatGPT integration "soon." You know the drill.
🔴 ChatGPT Adds A Trusted-Contact Suicide Hotline
New optional feature: nominate a friend or family member who gets notified if ChatGPT's safety system flags a self-harm conversation. Enrolled person consents. Reviewers escalate. Contact gets the call. The framing is careful, the underlying reality is bleaker — chatbots are now load-bearing in a mental health crisis, and OpenAI is building escalation paths for it.
🟠 Codex Now Drives Chrome. Sometimes.
Codex installs a Chrome extension and can take over tabs on macOS and Windows. Tested live on day-of-release: it loaded the page, highlighted tabs, then bailed with "direct page automation was blocked by an open extension UI." Eventually completed via search instead. Browser agents shipping in beta and labelled "stable." Same as it ever was.
🟣 OpenAI Also Shipped Pets In Codex. Yes, Pets.
Type `/pet` in Codex, get a tiny cloud robot that runs around your terminal. `/hatch` lets you generate custom pets via prompt. There's a community marketplace at codex-pets.net trading angry-Dario pets, Doom Guy pets, Alan Turing pets. The same company eliminating side quests in your workflow shipped side quests in their CLI. Self-aware, possibly. Funny, definitely.
🟢 Claude's Managed Agents Now "Dream"
Anthropic added Dreaming, multi-agent orchestration, and webhooks to Claude managed agents. Dreaming runs scheduled background passes over agent sessions, extracts patterns, restructures memory, surfaces recurring mistakes and shared team preferences. Not the chat-memory you're used to — this rewrites itself behind your back. The proactive-AI thesis, slowly becoming real. Inside enterprise tier only for now.
🔵 Claude Now Lives In Excel, Word, PowerPoint, Outlook
Side panel in all four. Crucially: context persists across apps. Reference the email from Outlook while writing the Word doc, then jump to Excel and Claude remembers everything. Microsoft built Copilot. Anthropic moved in next door. Office is now a two-AI building.
🟡 xAI Launches Grok 4.3. Cheap, Mid, Available.
Big jump from Grok 4.2 in the Artificial Analysis index. Still nowhere near Opus 4.7, GPT-5.5, or Gemini 3. The selling point: dramatically cheaper than the frontier — undercuts Anthropic and OpenAI on price-per-token. Grok 4 was already the discount aisle. 4.3 is the discount aisle with a bow on it.
🔴 Adobe Acrobat Becomes Notebook LM
New AI productivity agent inside Acrobat. Chat with PDFs, summarise, translate, and generate podcasts/presentations/blogs from the document. Tested it on the Attention Is All You Need paper. The podcast voices are noticeably worse than NotebookLM. The integration is noticeably better — it's where the PDFs already live.
🟠 Spotify Saves Your AI-Generated Podcasts To Your Library
New feature: agents like Codex, Claude Code, OpenClaw, and Hermes can now push generated audio briefings directly into your Spotify library. Sits next to your normal podcasts. Plays in the car. Daily AI news brief, delivered while you commute, voiced by something that probably sounds like a Notebook LM cosplayer. Personal podcast feeds, finally.
🟣 Apple's Camera AirPods Reach Advanced Testing
Bloomberg/Gurman: AirPods with built-in low-res cameras are deep into prototyping. Not for taking pictures. Not for filming. Just for feeding visual context to Siri so it knows what you're looking at. The Meta Ray-Bans without the part everyone actually liked. Vision-as-context, audio-only output. AirPods get eyes. You don't.
🟢 Your Next House Comes With A Data Centre In The Wall
Nvidia, PulteGroup, and startup Span partnered to install mini data centres inside the walls of new-build homes. Each unit: 16 Blackwell GPUs, 4 AMD Epyc CPUs, 3TB RAM. Runs AI inference on your home's spare electrical capacity. The pitch is local AI for the homeowner. The actual play is probably you renting unused capacity back to the hyperscalers, like the GPU version of feeding solar back to the grid. The future is your kitchen humming.
