AI Daily Digest — April 22, 2026 ⏱️ 8 min read

Today’s digest: Vercel got breached through a third-party AI tool and the attack chain is a wake-up call for every team that hands OAuth tokens to AI apps. Also: OpenAI flips the switch on pay-per-click ads in ChatGPT, and SoundHound swallows LivePerson.

🔥 Top Stories

Vercel Breached via Context AI Supply Chain Attack

This is the supply chain story everyone will be talking about this week. A Context.ai employee got hit with Lumma Stealer back in February. The attackers pivoted from stolen Google Workspace creds into Vercel’s infrastructure after a Vercel employee had granted “Allow All” OAuth permissions to Context’s AI Office Suite. Data — including API keys, deployment creds, and GitHub/npm tokens — is now listed on BreachForums for $2M. Vercel’s CEO says the attackers moved with “surprising velocity,” possibly AI-assisted. If you’re granting OAuth scopes to AI tools with your work account, maybe don’t. Read more →

OpenAI Turns On Cost-Per-Click Ads in ChatGPT

It’s official: ChatGPT now has CPC ads. Advertisers can bid between $3–$5 per click, down from the $60 CPM at launch ten weeks ago. OpenAI’s ad manager lets you optimize for views or clicks. The ad-supported AI assistant era is here, and honestly, the pricing feels aggressive for what’s essentially a conversational ad placement. Read more →

SoundHound AI Acquires LivePerson for $43M

SoundHound is buying LivePerson in an all-stock deal — $43M equity, $250M enterprise value after debt restructuring. The combined platform will serve 25 Fortune 100 customers across 30+ countries, merging SoundHound’s voice AI with LivePerson’s digital messaging that handles 1 billion messages per month. Stock dropped on the news, because markets gonna market. Read more →

Anthropic Now Requiring Government IDs for Some Users

Anthropic has started asking certain users for government-issued photo IDs and selfies to prevent access from US adversaries. This is new territory for an AI lab — identity verification at the model access layer. Whether this becomes industry standard or stays an Anthropic-specific policy remains to be seen. Read more →

AI-Generated Code Vulnerabilities Scaling 4x Faster Than Fixes

OX Security analyzed 216 million security findings across 250 orgs over 90 days. Raw alert volume grew 52% year-over-year, but prioritized critical risk grew nearly 400%. The “velocity gap” is real — we’re shipping AI-generated code faster than we can remediate the vulnerabilities it introduces. About time someone put numbers on this. Read more →

🛠️ New Tools & Releases

Tool What’s New Link
OpenAI GPT-5.4-Cyber Security-focused model for vulnerability detection; contributed to 3,000+ critical fixes via Trusted Access for Cyber program Details
Vercel Open Agents Open-sourced platform for building custom AI coding agents tailored to large codebases Details
GLM-5.1 (Zhipu AI) 744B MoE model, 40B active params, 200K context — fully MIT licensed. Legit open-source competitor. Details
Meta Muse Spark First model from Meta’s new Muse series, built by Meta Superintelligence Labs Details
Claude Code Routines Anthropic added scheduled automation to Claude Code — run coding tasks without active sessions Details

If you’re evaluating open models for production, check out our Best Open Model APIs for Developer Tools in 2026 comparison.

💰 Funding & Business

  • SoundHound + LivePerson: $43M all-stock acquisition creating an omnichannel conversational AI giant serving 30+ countries
  • AI coding tools market hits $12.8B in 2026 (up from $5.1B in 2024). GitHub Copilot holds ~37% share, but Cursor is gaining fast
  • India forms AIGEG: High-level inter-ministerial body to assess AI’s impact on employment and plan workforce transitions
  • ChatGPT ad CPMs already dropped from $60 to $25 in 10 weeks — the race to the bottom begins

📊 What Developers Are Discussing

  • OAuth supply chain risk: The Vercel breach has HN debating whether “Allow All” OAuth permissions for AI tools should be banned at the org level. Hot take: yes, obviously.
  • Vibe coding security: Research suggests 60–65% of AI-generated code contains vulnerabilities. The “just ship it” crowd is not happy about this stat.
  • 92% of orgs lack full visibility into AI identities running in their infrastructure — and most don’t even know what AI agents have access to what
  • MCP hits 97M installs: Anthropic’s Model Context Protocol is becoming the de facto standard. Every major AI provider now ships MCP-compatible tooling. See how the top AI coding agents compare
  • 84% developer adoption: Stack Overflow survey confirms AI coding tools are mainstream — the holdouts are shrinking fast

📝 Worth Reading

Similar Posts