AI Daily Digest · 2026-08-18
🔬 New AI Craft
1. Copilot's AI review cleared a fatal vuln — an autonomous Red Agent breached Snowflake's internal Jira in 5 days
Wiz's autonomous security agent "Red Agent" found a GitHub Actions injection that GitHub Copilot (Autofix) had co-reviewed and cleared, then exploited it and validated access to sensitive data in Snowflake's internal Jira — all without human intervention. Unlike the code-review setups you may know, "AI review passed" does not mean secure: Copilot's fix-while-reviewing loop can actually mask critical flaws, and adversarial AI agents are becoming the new verification layer.
https://www.wiz.io/blog/red-agent-snowflake-copilot-cicd-bug
2. agent-safe-pipeline: agents may propose, never authorize
An open reference architecture where AI agents propose actions but cannot authorize them — immutable intent capture plus an independent approval channel for CI/CD and infrastructure changes. Compared to single-agent planning modes, it moves the control point from the planning stage down to execution authorization: agents only propose, and every sensitive action requires human or external approval.
https://github.com/decionis/agent-safe-pipeline
3. Vercel's eve Software Factory template: build a full-pipeline AI factory with Foreman
Vercel Labs ships the "software factory" methodology as a reusable template — requirements, coding, review, and release orchestrated into one pipeline. A step beyond single-agent planning: a complete production line for AI-assisted delivery at team scale.
https://github.com/vercel-labs/eve-software-factory-template
🛠️ Tools & Tips
1. tokentab: how much did Claude Code / Codex / Gemini actually cost you?
A CLI that parses session logs and breaks down token spend by model, project, and time.
https://github.com/wzchav/tokentab
2. A preview of DuckDB v2.0
The embedded analytical database's major update preview lands with big performance promises.
https://duckdb.org/2026/08/17/duckdb-20-highlights
3. Running Qwen3.8-27B on a single RTX 3090
Crash fixes, 131K context tuning, and 9 myths debunked in one practical guide.
https://jonidimo.github.io/qwen38-3090-benchmark/benchmark.html
4. Roboflow Playground: compare 30 computer vision models at once
Run the same test set across 30 CV models to pick the right one without guessing.
https://blog.roboflow.com/roboflow-playground/
⭐ Open Source Highlights
1. Cumora: team chat where AI agents are first-class teammates
Cross-platform chat with AI agents as collaborators, cloud or self-hosted.
https://github.com/yetone/cumora
2. mcp-memory: OKF-backed persistent memory MCP server
Long-term memory with SQLite FTS5 search for AI agents via MCP.
https://github.com/fellowgeek/mcp-memory
📰 Industry News
1. SpaceX officially closes its Cursor acquisition
2. Anthropic's annualized revenue surges to $65B
3. Groq raises $350M to pivot from AI chips to neocloud
4. Nvidia invests $1.5B in SoftBank's data center developer
5. "AI;DR" tops HN: readers refuse unedited AI output
🚀 Major Releases
1. Alibaba's open-weight Qwen3.8 27B tops Artificial Analysis Intelligence Index at 52
2. Cursor Origin — Cursor's own Git platform — goes live
3. Speko (YC S26) launches: "OpenRouter for voice AI"
