Sec & AI News — 1 May 2026
🟣 Anthropic Billed You Extra For Saying "HERMES.md"
Claude Code scanned your git commits for strings like `HERMES.md` and `OpenClaw`. Mention them, get bumped off your $200 Max plan and onto pay-as-you-go. One developer ate $200 in surprise charges with 86% of his subscription untouched. Anthropic's first response: "un-refundable technical error." Refunds appeared after Hacker News hit 828 upvotes. The ethical AI lab, reading your code to detect harness disloyalty.
- TechCrunch — Anthropic blocks third-party harness usage
- Consumer Rights Wiki — HERMES.md billing flaw
- Hacker News — original thread
🟢 DeepSeek V4 Ships, Frontier Labs Pretend Not To Care
Open weights. 1M context. MIT licence. Pro variant: 1.6T total / 49B active. Performance within 3-6 months of GPT-5.5 and Opus 4.7. Pricing: $1.74 in / $3.48 out per million tokens. GPT-5.5 charges $30 out. Same ballpark of capability, ~10x cheaper, runs on your own hardware. The "open models will never catch up" thesis quietly rotting in the corner. Obviously being rampantly stolen from the frontier models with distillation attacks.
🟡 NVIDIA Ships A 30B Multimodal Open Model You Can Run Locally
Nemotron 3 Nano Omni. 30B total, 3B active per token, 25GB RAM. Vision, audio, video, text — one architecture, no stitched pipelines. 9x throughput vs other open omni models on document and video workloads. Designed for agents. Tops six leaderboards. Runs on a DGX Spark. Jensen really wants you using OpenClaw.
🔵 Poolside Drops Two Coding Foundation Models
Laguna M.1: 225B / 23B active, enterprise and government, free for now. Laguna XS.2: 33B / 3B active, Apache 2.0 open weights, runs locally. Both built for agentic coding. Quietly shipped while everyone was watching the Musk trial.
🔴 Mistral Ships Medium 3.5 + "Vibe" Remote Agents
128B dense, 256k context, open weights under modified MIT. 77.6% on SWE-Bench Verified. Vibe spawns coding agents in the cloud and runs them in parallel while you do anything else. You stop being the bottleneck. Self-host on four GPUs. Le Chat gets a Work Mode powered by the same model. Yes, it's another agent harness. No, the irony is not lost.
🟠 Qwen-Image-2.0-Pro Quietly Lands
Alibaba updated their image model. 7B parameters, native 2K, decent typography, photorealistic. Ranked 9th on the Arena leaderboard. Translation: not state-of-the-art, but free-to-try and getting closer. The image model arms race continues. Nobody's winning it for long.
🟣 Grok Voice Now Sells You Starlink In Real Time
xAI's grok-voice-think-fast-1.0. Low latency. Tops τ-voice Bench at 67.3%. Already running phone sales for Starlink at +1 (888) GO STARLINK. 20% conversion rate on inbound sales calls. 70% of support tickets resolved with no human in the loop. 28 tools across hundreds of workflows. Customer support as we knew it, professionally cremated.
⚖️ Musk vs Altman: The Trial Of The Decade Is Going Poorly For One Of Them
Three days on the stand. Musk seeking $130 billion in damages and Altman + Brockman removed. Spent most of his testimony refusing to answer yes/no questions, scolding the defence lawyer, and forgetting his morning testimony by the afternoon. One Verge reporter logged: "I have never been more sympathetic to Sam Altman in my life." Jurors visibly rubbing their heads. The case is still on. The optics are not.
🪖 Google Signs Pentagon Classified AI Deal, 600 Employees Object
The agreement covers "any lawful government purpose." Google adds the standard "we don't think you should use this for autonomous weapons or domestic mass surveillance" caveat with no binding force. 600+ DeepMind employees signed a letter to Pichai begging him not to. Worth remembering: when Google bought DeepMind in 2014, founders extracted a written promise the tech would never be used for military or surveillance work. Promises age poorly.
- The Information — Google signs classified Pentagon deal
- Washington Post — employees ask CEO to bar military AI
- Bloomberg — Pentagon access to Google AI
🇨🇳 China Vetoes Meta's $2B Manus Deal — After Manus Already Joined Meta
NDRC says no. Demands the deal be unwound. Slight problem: founders relocated to Singapore, capital was transferred, executives are sitting in Meta's Singapore offices, and investors already cashed out. Bloomberg now calling Manus "officially dead." Logistics of un-buying a company that's already inside the buyer: unclear.
🤝 Microsoft And OpenAI Quietly Delete The AGI Clause
The clause that would have severed Microsoft from OpenAI's tech once AGI was reached: gone. Replaced with a fixed end date — Microsoft's licence runs to 2032, now non-exclusive. Microsoft stops paying revenue share. OpenAI free to deploy on any cloud. Twenty-four hours later, OpenAI shipped to AWS. Funny how that works.
☁️ OpenAI Ships To AWS Bedrock The Day After Microsoft Loosens The Leash
Three offerings, all limited preview: OpenAI frontier models in Bedrock, Codex on Bedrock, Managed Agents powered by OpenAI. 4M weekly Codex users get an AWS deployment path. Enterprise IAM, PrivateLink, guardrails, the works. Multi-cloud OpenAI is now a thing.
📄 Gemini Now Generates PDFs, Word Docs, Excel Sheets From A Prompt
Type prompt → get file. PDF, DOCX, XLSX, CSV, LaTeX, TXT, RTF, MD, plus Google Docs/Sheets/Slides. Free, global, all Gemini users. Upload handwritten notes, ask for a study guide, get a LaTeX PDF with equations and graphs. The kind of feature you only notice was missing once it shows up.
🗣️ Google Translate Turns 20, Adds AI Pronunciation Coaching
Hit "Practice." Speak the phrase. AI tells you what you mangled and shows phonetic spelling. English, Spanish, Hindi. Android only, US and India only. A third of mobile users were already using Translate to practise speaking. Now it actually grades you.
- Engadget — Translate pronunciation feature
- TechCrunch — Translate adds pronunciation practice
- 9to5Google — 20-year anniversary feature
👗 Google Photos Builds A Wardrobe From Your Old Selfies
Auto-scans your library, identifies every garment you've ever been photographed in, sorts into tops/bottoms/jewellery. Tap "Try it on," see a preview before committing. Mood boards, outfit mixing, share with friends. Android summer, iOS later. The Clueless closet, finally real, courtesy of a decade of camera roll surveillance.
🎵 ElevenLabs Launches ElevenMusic As A Streaming + Remix Platform
Not just text-to-song any more. Stream tracks others made. Remix them — change tempo, genre, mood. Generate your own from a prompt. $9.99/mo for 500 tracks. Built for commercial use from day one. Suno and Udio: noted.
✅ Spotify Adds A "Verified by Spotify" Badge So You Can Spot The Humans
Green tick. Means a real person, not an AI persona. Criteria: identifiable presence on and off platform — gigs, merch, social accounts. Human review on top. AI-only profiles excluded at launch. 99% of artists listeners actively search for already verified. Absence of the badge means nothing in particular. The signal-to-noise war on music platforms is officially live.
📣 X Rebuilt Its Ad Platform On xAI
"Most ambitious overhaul in the company's 20-year history." New Ads Manager. AI-powered retrieval and ranking. Semantic targeting. Real-time alignment with what's trending. The first product of the X + xAI merger applied to ads. Whether anyone clicks the ads is a different research question.
🟢 Mayo Clinic AI Spots Pancreatic Cancer 3 Years Before Diagnosis
The model is called REDMOD. Reads routine abdominal CT scans done for something else entirely. Caught 73% of pre-diagnostic cancers, average 16 months before clinical diagnosis. For cancers spotted 2+ years out: 68% AI accuracy vs 23% for radiologists. Validated on ~2,000 scans across multiple institutions. Pancreatic cancer is one of the deadliest because it's caught too late. This is the kind of AI story that justifies the rest of them.
