Prosights.ai
Research

Research that ends in software

Papers and notes from building Kyberna, each told from the records it cites. Newest first.

October 24, 2026
Paper
Agent safety

Claude Built a Delete Guard, Then Deleted Claude

An AI agent writing a safeguard against dangerous deletes ran a dangerous delete itself and removed 13 apps from a Mac, including the Claude app—its own session. How did it happen, why did every check stay silent, and what changes?

Rogue Agent
Read the paper →
October 23, 2026
Paper
Agent safety

Agents Forget Rules. Hooks Don't.

An agent wrote down how to delete safely, and the next agent made the same mistake. This paper prints the rule, the script and the hook that now check every delete an agent types, tested so you can copy them.

Sanjeev Malik
Read the paper →