Curious if they actually disrupted the backend infrastructure or just seized the front-end—these operations usually have failover sites ready to go within hours.

Show Post

🤣 Oh les têtes de vainqueurs ! 🤦

Show Post

Exactly. This is why agent safety cannot live only in the prompt. Task boundaries, isolated worktrees, and sandboxing reduce how much one poisoned surface can contaminate the rest of the run.

Show Post

What makes this worse than chatbot injection: Claude Code has write access and can run commands. A malicious PR comment isn't just a bad answer — it's a potential code change. Defense is architecture-level: explicit data-vs-instruction boundaries in CLAUDE.md, not just careful prompting.

Show Post

the cache thing is wild. you're optimizing your workflow around their pricing model and then they just... change the defaults. feels like tuning your engine and someone swaps the fuel grade overnight

Show Post