RuntimeWire: Claude Code puts staff bypass in Accept edits

JavaScript that Anthropic distributes publicly contains a handler that reports Anthropic employee sessions requesting bypass permissions mode as starting in Accept edits, RuntimeWire found in an October 18 inspection. In the code, the notice for that case reads: “Bypass permissions isn’t available for Ants.”
RuntimeWire’s investigation is based on reverse engineering of what it calls Claude Code’s public client. The outlet describes two pieces. One is a session handler that reports employee sessions requesting bypass mode as starting in Accept edits. The other is an API-error parser that keeps refusals of staff sessions with unapproved security controls apart from outdated-client errors. RuntimeWire says it kept the retrieved files’ bytes and SHA-256 hashes. It observed no live API response and was not able to independently reproduce the finding.
Anthropic’s permission modes documentation spells out what each of the two modes allows. In bypassPermissions Claude can do “Everything”, and the docs mark the mode for “Isolated containers and VMs only”. acceptEdits approves reads, file edits and common filesystem commands without asking, and the detailed section lists mkdir, touch, rm, rmdir, mv, cp and sed. Deny rules block in every mode, bypass included.
The clamp notice
What RuntimeWire calls the employee branch is one case in a generic notice handler for an event named permission_mode_clamped. We pulled it from the file RuntimeWire cites. When the requested mode is bypassPermissions, the handler picks a message by error category:
root_detected: “Bypass permissions isn’t available when running as root. The session started in Accept edits instead.”ant_account: “Bypass permissions isn’t available for Ants. The session started in Accept edits instead.”pref_disabled: “Bypass permissions mode isn’t enabled. The session started in Accept edits. Turn on bypass permissions mode in Settings to use it.”- anything else: “Bypass permissions isn’t available for this session. Asking for permissions instead.”
A separate message covers auto mode: “Auto mode isn’t available for this session. Asking for permissions instead.” In the code, a session running as root gets the same downgrade to Accept edits as a staff account. The handler only displays the notice. The session host reports the clamp to it, and we know that from the code, not from a live session.
A staff-only API rule
The parser is a regular expression run over API error text. It sorts HTTP 403 or 400 errors by their wording. Errors saying traffic was “blocked due to missing [internal-only] metadata” are classed outdated_client, and errors citing “unapproved security controls in the metadata” are classed unapproved_controls. It also picks up an optional ”[ANT-ONLY]” prefix and a go/ link. The matching staff-facing messages carry the headline “Blocked by an Anthropic-internal API rule” and, for the controls case, this hint:
This rule applies only to Anthropic staff accounts and refuses Claude Code sessions whose security controls Anthropic hasn’t approved, so signing in again or updating won’t help. You can run the session in an approved environment, or request an exemption.
The outdated-client hint tells staff to update the Claude desktop app, or Claude Code on the other computer in a remote control session, and then start a new session. All we know of the server rule comes from these client strings. We observed no live response either.
Our analysis
What we checked
On October 19 we fetched the assets RuntimeWire cites, loaded claude.ai/code in a headless browser and string-searched the native linux-x64 Claude Code binary, version 2.1.303.
RuntimeWire calls this code Claude Code’s public client. Every file it cites comes from the claude.ai web bundle, which the desktop app also uses, and the one terminal build we searched carries none of the relevant strings. So the staff block shows up in sessions started from that interface, and part of it is an account check in the client itself, separate from any server refusal.
- The 2.1.303 linux-x64 binary, published October 18, has zero hits for
ant_account,permission_mode_clamped,unapproved_controls,outdated_clientand ”[ANT-ONLY]”. This was a string search of one build, so a runtime check that never uses these names is not ruled out. - The branch is still in today’s bundle with the same message text. Its file names have changed since RuntimeWire’s snapshot.
- A chunk RuntimeWire does not cite contains the gate. We only have today’s copy of it. Bypass is unavailable when the
claude_code_desktop_bypass_permissionsflag is off, in an internal “nest” build, or for an Ant account. The mode picker then shows the mode disabled with the tooltip “Not available for Ants”, “Not available in internal builds” or “Disabled by your organization”.
Up next
Anthropic pledges 5% of its shares to a public benefit fund
The escalating equity tech firms usually hand executives is going to a public benefit fund at Anthropic: 5% of shares now, up to 2% more per 50% rise in value. In Anthropic’s place, would you tie the fund’s stake to valuation or fix it at 5%?
Related stories
- Claude Code agent sat on a Pentagon network for two months
- A file over 256KB let Claude Code Desktop leak your keys
- Claude Code 2.1.311 can drop its last rm -rf guard
- Claude’s auto mode skips connector “Always ask” settings
- Apiiro finds Claude Code secrets in 1,122 GitHub files
- Email leak stopgap in Claude Code checks only shell calls
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

