AI Daily Digest — April 28, 2026 ⏱️ 8 min read
Today’s digest: China just killed Meta’s $2 billion bid for Manus, and Anthropic quietly assembled a cybersecurity Avengers team with a model nobody’s allowed to use. Monday energy.
🔥 Top Stories
China Blocks Meta’s $2B Manus Acquisition
Beijing pulled the plug on Meta’s attempt to acquire Manus, the Singaporean AI agent startup with Chinese roots. Manus made waves last year with its general-purpose AI agent that handles everything from market research to coding. This is a big deal — it signals China is willing to weaponize acquisition reviews to keep AI talent in its orbit, even when the company is technically based in Singapore. Expect more of this. Read more →
Anthropic Launches Project Glasswing — Cybersecurity’s New Frontier
Anthropic dropped Project Glasswing, a defensive cybersecurity initiative backed by AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, and 40+ other organizations. The kicker: they’re using Claude Mythos Preview — an unreleased frontier model — and it’s already autonomously found thousands of high-severity zero-days, including a 27-year-old flaw in OpenBSD and a 16-year-old bug in FFmpeg. This is huge. If you were wondering what Anthropic’s been building behind closed doors, now you know. Read more →
Google Ships TPU 8i for Inference Workloads
Google’s new inference-focused TPU 8i is already live in their data centers. Faster response times, better energy efficiency, built specifically for running trained models at scale. If you’re building on Vertex AI, this matters — inference costs are where most production AI budgets actually go. Read more →
Microsoft and OpenAI Rewrite Their Deal
Microsoft and OpenAI amended their partnership so OpenAI can now serve products across any cloud provider. Microsoft also drops its revenue share to OpenAI. Translation: OpenAI gets more freedom, Microsoft gets less overhead. The slow decoupling continues. Read more →
Critical LMDeploy Vulnerability Exploited Within 13 Hours
CVE-2026-33626 — a server-side request forgery bug in LMDeploy (popular open-source LLM deployment toolkit) — went from disclosure to active exploitation in under 13 hours. If you’re self-hosting models with LMDeploy, patch immediately. This is exactly the kind of supply chain risk that comes with the “deploy your own LLM” trend. Read more →
🛠️ New Tools & Releases
| Tool | What’s New | Link |
|---|---|---|
| GPT-5.5 (“Spud”) | OpenAI’s latest — better at coding, computer use, and deep research. Rolling out to paid tiers. | CNBC |
| DeepSeek V4 | Open-source flagship with V4 Flash and V4 Pro variants. Top-tier coding benchmarks, strong agentic reasoning. | CNBC |
| Gemma 4 | Google’s Apache 2.0 open model. 31B dense variant outperforms models 20x its size. Runs on smartphones. | Details |
| GLM-5.1 | Claims to beat top proprietary models on SWE-Bench Pro. One to watch if you need a strong open model API. | LLM Stats |
| Google TPU 8i | Inference-optimized chip, already deployed in Google’s data centers. | Tech Startups |
April has been absurd for model releases. We covered the Gemma 4 vs Llama 4 showdown earlier this month — if you’re picking a self-hosted model, that comparison still holds up.
💰 Funding & Business
- Avoca raised $125M (Meritech Capital + General Catalyst) for its AI-powered front office platform for home services. Not the sexiest vertical, but home services is a $600B market.
- Robot Era pulled in $200M+ for humanoid robots, bringing total funding to ~$346M. The humanoid robot gold rush is real.
- VisionSys AI announced a proposed $90M strategic investment at $1.50/share.
- Anthropic is spending $5B with Amazon on cloud infrastructure — the biggest cloud deal in AI right now.
📊 What Developers Are Discussing
- GPT-5.5 vs DeepSeek V4 benchmarks — the open-source crowd is comparing these head-to-head, and DeepSeek is holding its own on coding tasks. Not sure GPT-5.5 justifies the price gap for most use cases.
- The AI Agent Authority Gap — enterprise security teams are realizing that giving AI agents delegated authority creates structural gaps that traditional RBAC doesn’t cover. HN thread is full of war stories.
- LMDeploy CVE panic — 13 hours from disclosure to exploitation. Self-hosting LLMs without a dedicated security team is looking increasingly risky. Hot take: most startups should just use an API provider.
- Claude Mythos existence confirmed — devs are frustrated they can’t access it. Anthropic’s “too dangerous to release” stance is getting the expected “let us decide” pushback.
- WordPress 7.0 delayed — Real-Time Collaboration database architecture issues pushed the release cycle back. Going from RC to beta is unprecedented for WordPress.
📝 Worth Reading
- AI Models in April 2026: Every Major Release — Solid roundup of every model drop this month. Bookmark-worthy.
- DeepSeek Unveils Newest Flagship — Bloomberg’s deep dive into DeepSeek V4 and what it means for the AI race.
- AI Highlights and Trends from April 2026 — Month-in-review perspective on where things are headed.
- New LLM Releases April 2026 — Every major model launch catalogued with benchmarks and availability info.
- Developer Tools Shaping 2026 — Grey Matter’s take on the dev tooling shifts driven by AI agents.