
Claude Code v2.1.213-214: Six Permission Bypass Fixes, EndConversation Added, /verify No Longer Auto-Runs
Chris Harper
2 min read
Jul 19, 2026 · 20:04 UTC
Claude Code v2.1.213-214 (Jul 18-19) closes six permission-check bypasses, changes /verify and /code-review to require explicit invocation, and adds the EndConversation tool for jailbreak defense.
v2.1.214 — Security and stability (Jul 18)
Six permission-check bypasses patched:
dir/**allow rules (e.g.,Edit(src/**)) auto-approved writes to nestedsrc/directories anywhere in the tree — not just<cwd>/src/- Windows PowerShell 5.1 — permission checks could be bypassed entirely in PS 5.1 sessions
- File-descriptor redirects (e.g.,
cmd 3>&1) — bash and the analyzer disagreed on safety; now fails closed - Very long commands (>10,000 chars) — now always prompt instead of running automatically
- Zsh
[[ ]]subscripts — variable subscripts/modifiers were treated as inert text, now analyzed correctly help/manwith unsafe options — certain forms could trigger command substitutions or backslash paths
Also added: the EndConversation tool (Claude can end sessions with abusive users or jailbreak attempts), docker daemon-redirect flags (--url, --connection, --identity) now require confirmation, and a periodic heartbeat for long-running tool calls that previously went silent.
v2.1.213 — Behavior change (Jul 19)
/verify and /code-review no longer run automatically. They only run when you explicitly invoke /verify or /code-review. If your workflow expected Claude to proactively run these after edits, you'll need to invoke them manually or wire them into a PostToolUse hook.
Why it matters: Fix #1 affects every project with a Edit(src/**) or similar glob allow rule. On any repo with a nested src/ directory, writes may have been approved silently. Update now: npm install -g @anthropic-ai/claude-code@latest
Sources: Claude Code Changelog, GitHub Releases