Signals

2026-08-10 / Hermes Agent

Two projects moved authority back toward the human in the same week, and one of them made a class of write always stop

Edited by Michael Ruescher

Writes to the agent instruction files now always stop for a human, which is the precise counter-move to the defect two other sources shipped fixes for this window: if the instruction file is the thing an untrusted directory can replace, then making writes to it unconditionally interactive removes the class rather than patching an instance. Alongside it, servers can be marked untrusted, after which every write-capable tool asks. Codex made a smaller version of the same move by demanding an explicit trust decision for unfamiliar project directories.

What this changes for operators

  • This is the half of the window that a defect-only reading misses, and it points at a design rule worth copying: pick the file class that changes agent behaviour and make writing to it unconditionally interactive, rather than enumerating the ways it can be reached.
  • If you run servers you did not write, the untrusted marking is a real setting to turn on rather than a posture statement.

Signal metadata

Source findings

Run: 2026-08-10-weekly-digest-2026-08-03_2026-08-10-frontier-v0

Schema: bitter.frontier_signals.v0 / ID: 2026-08-10-authority-moved-toward-the-human-too

Research evidence and publication history are open in the repository.