Under the hoodTier: how much it should matter to you
1Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
Prompt SafetyArea: what it touches
Internal ChangesKind: in v2.1.318,
Internal ChangesSection of the release
What
Claude Code removes text that pretends to be its own internal control tags, so outside content cannot pass itself off as instructions. That clean-up can now be preceded by a quicker first check, switched by a remote setting. Claude Code also counts how often the check matches, misses or fails.
When the server sends nothing, the setting falls back to not treating the check as safe to use.
Why
This is about making the defence against prompt injection, where text tries to trick Claude into following it, faster without weakening it.
