Claude Code v2.1.321 buried an MCP OAuth token leak and three sandbox gaps in a very long changelog
Most Claude Code release notes are a blur of quality-of-life fixes. v2.1.321, published on 9 November, is different: buried in its very long changelog are several fixes that touch the exact boundaries I rely on when I let an agent loose in a repo. I read the whole thing so you only have to read this.
The MCP OAuth fix
The one that made me sit up: the release notes say it fixed "MCP OAuth requests sending secrets such as refresh tokens to another host when a server redirected them". In plain terms, if an MCP server answered an OAuth request with a redirect, secrets like refresh tokens could end up going to a different host than the one you signed in to. The notes don't give a CVE or advisory ID, and I won't invent a severity for it. But the failure class is a familiar one: credentials should never follow a redirect to another origin.
If you connect remote MCP servers that use OAuth, update, and consider the tokens you had issued before today as worth rotating if the server is one you don't fully control.
Sandbox and permission-check fixes
Four items in the same release concern Bash permissions:
- A read-only Bash command could run outside the sandbox without a permission prompt.
- Bash permission checks could judge a file write from the wrong folder and skip the prompt.
- Deny rules and protected-file prompts did not apply to the file that a
cpormvinto a directory writes. cpandmvspelled by path, such as/bin/cp, now get the same path checks as the bare commands.
The cp/mv one is the most interesting to me. cp a.txt some/dir/ never names the destination file; it is derived from the directory. A rule matcher that only looks at literal arguments will happily miss that. My deny rules for .env and similar files assumed they covered every write path, and this is a good reminder that a deny list is only as good as the command parsing under it.
Smaller things worth knowing
A few other changes affect day-to-day use. claude remote-control gained --settings <file>, which applies a settings file to every session it starts. Remote Control sessions also keep the folder they were started in, so a connected client can no longer move a session to another folder. On the model side, the notes list Claude Fable 5.2 (claude-fable-5-2) as the new default Fable model.
Note one behaviour change that cuts the other way: mods Claude writes now turn on without asking in auto and bypass modes, and mods' own tool calls raise no permission prompt, though deny rules and organization policy still apply. That is a deliberate convenience trade-off, and worth knowing if you run those modes.
What I'm doing about it
I updated before writing this. Then I reviewed my deny rules with one question in mind: which of them assumed a write has to name its target file? Anything that relies on pattern-matching shell commands deserves the same audit. The sandbox is the real backstop, and a "read-only command escaped it" fix is a reminder that the backstop has bugs too.
None of this is dramatic on its own, and the notes are honest about what was fixed without hyping it. But if you run your agent unattended, a changelog is a security document. Read it like one.