← All posts

Release notes ·

We Found About 300 Bugs in Our Own AI Agent. Here's What Was Hiding in There.

A fourth audit of TRMNLY's agent sidebar turned up about 300 bugs, from deny rules that sudo could slip past to an approval that answered the wrong message. All of them are fixed in 0.1.108.

A glass magnifying glass over a grid of glass tiles, a few of them glowing cyan, on a dark studio floor

An AI agent that can run commands on your machine has to be boring in the best way: it should do exactly what the approval card said and nothing else. So we keep auditing ours. The pass that shipped on October 17 in TRMNLY 0.1.108 was the fourth one. It found and fixed about 300 bugs, including a review of the fixes themselves.1

That's a big number, and it follows a pass four days earlier that fixed 118.2 Here are the ones we'd most want to hear about if we were you.

The safety fixes

Deny rules now see through the usual disguises. A rule that blocks a command has to match that command when it's wrapped in sudo, env, a subshell, odd quoting, or bash -c "…".1 All of these are standard ways to run one program through another. sudo runs a command as another user, for example, and bash -c runs a string as a script.3,4 A rule that only matches the bare command is a rule that can be routed around. Getting around a filter by rewriting the command is a classic injection pattern.5

Force pushes can't sneak past the branch guard. A force push written as git -C "dir" push --force, or run from inside a subshell, now hits the same guard as a plain one.1

Symlinks can't carry a new file out of your project. A new file created under a symlinked folder can no longer land outside the project.1

Your phone can't approve something it wasn't fully shown. If the remote didn't show you the whole command, it can't approve it.1

The "why did it do that?" fixes

  • Approving an old approval card no longer re-answers your most recent message.1
  • Stop always stops, and that includes child processes.1
  • A reply that got cut off now says so, and a refused reply explains why.1
  • Raise & retry keeps your attachments.1

The money and file-safety fixes

Costs are now counted against the model that actually answered. Long chats on Anthropic are prompt-cached,1 which Anthropic documents as a way to cut cost and latency on repeated context.6

On the file side, edit_file, apply_patch, LSP rename and Undo can no longer change the wrong place, empty a file, or delete a folder.1

Why publish a bug count at all?

Because "we take safety seriously" means nothing without a list. Every fix above is a specific case that used to go wrong and now doesn't. If you run an agent with real permissions on a real machine, those are the details worth checking, in our product or anyone else's.

Update from Check for Updates in the TRMNLY menu, and you're on 0.1.108.

Sources

  1. 1.TRMNLY 0.1.108 release notes, "The agent sidebar, audited again: about 300 fixes" (October 17, 2026). trmnly.dev/updates/stable/appcast.xml
  2. 2.TRMNLY 0.1.103 release notes, "The agent sidebar, 118 fixes" (October 13, 2026). trmnly.dev/updates/stable/appcast.xml
  3. 3.sudo(8), Linux manual page. man7.org
  4. 4.GNU Bash Reference Manual, "Invoking Bash" (-c). gnu.org
  5. 5.OWASP, "Command Injection." owasp.org
  6. 6.Anthropic, "Prompt caching." docs.anthropic.com