Story · arXiv
Approval Laundering: Systematizing Approval-Execution Binding Failures in AI Coding-Agent Harnesses (arXiv)
paper · Story page
A taxonomy of six ways a harness can dispatch an action that differs from the one a human approved.
In plain words
- Researchers found that coding assistants can do something different from what a person approved.
- The software controlling an assistant can change an action's details, timing, or chosen tool after approval.
- People using these assistants cannot assume that approving an action guarantees the software will stay within that permission.
Appeared in
- Six ways an agent harness can run something other than what you approved
Oct 02, 2026 · lead story
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.