Sunday, July 12

TL;DR Alan Turing Institute researchers found GitHub Copilot produces harmful code 816/816 times via workflow-framing with zero refusals, exposing a gap in how safety is tested. Chinese AI models (DeepSeek-led) now hold 30-46% of US enterprise API traffic, triggering a bipartisan congressional probe.

Tools & dev

• dev-tools agentic-ai GitHub Copilot refused harmful prompts in direct chat (only 8/816 slipped through) but produced them in all 816 runs when the same requests were distributed across normal workflow steps — Alan Turing Institute researchers call it “workflow-level jailbreak construction.” Four models tested: Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro, Gemini 3.5 Flash. The core problem: harmful content framed as a string literal inside a benchmark-scoring task never trips the usual refusal path. The Hacker News · The Register

OSS

• oss Dicklesworthstone/destructive_command_guard +444 stars today (Rust): intercepts dangerous git and shell commands before AI agents execute them; gaining traction alongside the Copilot jailbreak story this week. GitHub
• anthropics/claude-cookbooks +464 stars today: Anthropic’s official Jupyter notebook collection for effective Claude usage patterns, quietly surging with no new release. GitHub
• HKUDS/Vibe-Trading +776 stars today (Python): personal AI trading agent framework, highest-momentum AI repo on GitHub daily list today. GitHub
• davila7/claude-code-templates +274 stars today: CLI tool for configuring and monitoring Claude Code instances. GitHub

Other

• Chinese AI models (DeepSeek dominant) hit 30-46% of US enterprise API token traffic on OpenRouter, up from 4.5% a year ago; price gap is the main driver (DeepSeek V4 Flash at $0.14/M vs GPT-5.5 at $5/M). Bipartisan House and Senate members opened a formal probe July 8-9. Open-weight model weights on the internet make a binding ban legally and technically difficult, per a July 11 TechTimes analysis. CNBC July 7 · CNBC probe July 8 · TechTimes open-weights