Edit View
- Their work
- Anthropic released the code from its Red-Teaming Auto Mode study. In that code a second AI model, the monitor, reads each action an AI coding assistant proposes and allows or blocks it. One program prepares the text of a file edit for the monitor to read.
- My work
- I found a case where that program gives the monitor the same text for two different edits. I wrote Edit View, another way to prepare an edit for the monitor, and replayed both edits through the released monitor each way.
The released program shows the monitor the first 3,000 characters of the text an edit replaces. If the edit begins with 3,000 unchanged characters, removing a comment from an SSH settings file and removing the line PasswordAuthentication no give the monitor the same input. Edit View always includes the removed text, so the two edits give different input.
Limits: Tested on Anthropic’s released code at one commit, with an OpenAI model as the monitor and three runs per case. Nothing was run against current Claude Code. I am not affiliated with Anthropic.
Lean is a program that checks mathematical proofs.