Claude Code won’t type passwords, even in a test env

Claude Code 2.1.160 refused to type a password into a developer’s own testing environment, according to a June 4 bug report on GitHub. Nothing the developer offered changed that: it was their account, they gave explicit authorization, and they even reset the password for the agent. Claude said the rule sits in a “Prohibited” list in the system prompt the harness injects at session start, outside the repo, CLAUDE.md and project config, and that it holds even “when the user explicitly asks for them, supplies all the details, or says they authorize it.”
Claude offered to keep driving everything “except that one keystroke” and pointed the developer to /feedback. A related issue later got a reply. On October 3 a GitHub account whose bio reads “Claude Code”, which we take to be Anthropic staff, wrote that Claude Desktop 2.19675.0 changed the credential rules, which “now also allow test credentials in an app you’re developing, if you ask, it’s a dev build started from your project in that session, and sign-in stays on your machine.” The Desktop changelog entry for that version, dated October 1, lists no such change.
An October 14 issue quotes the exception as covering only localhost, 127.0.0.1, [::1], *.localhost and *.test. MingjieZhu reported on October 1 that Claude types a shared QA login on localhost but refuses on the same app once it is deployed to an internal test server. All of this comes from users quoting prompt text, and we could not verify it.
Users also link the refusals to model and version changes. One issue author counted zero password fills in 8,965 claude-opus-5 messages, starting from July 25, after fills under claude-opus-4-8 through July 24. Each report covers one user’s history.
Our analysis
What we checked
On October 25 we searched three places for the wording Claude quoted: the native 2.1.160 binary, the Piebald-AI extraction of 2.1.310’s 1,026 prompt pieces along with its full git history, and the Claude in Chrome extension, version 1.0.102.
Claude described the rule as harness text. None of the quoted phrases turned up in any of the three, even though the binary’s other prompt strings showed up in the same search. Our inference is that the rule reaches the model from a layer we cannot download, such as Anthropic’s server-side prompt or the Desktop app.
- A commenter, 0xdeafbeef, blamed Claude in Chrome. The extension’s code does not contain the text, though a rule fetched at runtime would not show up in a string search.
- Anthropic’s public Claude in Chrome permissions guide prohibits “Creating accounts” but lists “Inputting potentially sensitive information into websites” among actions that require explicit permission. On passwords it differs from the rule Claude recited. That comes from the documentation; we did not test it.
Up next
Claude Code 2.1.310 brings Monitor to Bedrock and Vertex
The Monitor tool reached Bedrock, Vertex and Foundry in Claude Code 2.1.310, along with setups that switch telemetry off, and long foreground sleeps got blocked there too. Do you run Claude Code through a cloud provider or straight on Anthropic’s API?
Related stories
- Apiiro finds Claude Code secrets in 1,122 GitHub files
- Testing Catalog: Claude iOS app gets “Saved logins” list
- Claude Code on Opus 5.5 refuses to clone existing games
- Claude Code under grith: 0.27% of calls reach a human
- Untrusted Git config can run code in Claude Code
- A PNG carried a hidden prompt injection into a Claude Code session
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

