Claude Code v2.1.223, released August 6, 2026 at 00:52 UTC with 19 changes, patches three distinct Bash permission-bypass classes. Each let crafted commands hide portions of themselves from the approval dialog -- meaning you could approve one thing while Claude executed something else. Here's exactly what each exploit class was, how it worked, and what the fix actually does.
I run Claude Code in agentic mode across multiple production workflows. Knowing what the permission checker catches -- and what it was silently missing until today -- matters more than most people realize when you're running agents with Bash access. This release is the fourth consecutive August 2026 update to include permission-layer security fixes.
What is a permission bypass in Claude Code?
A Claude Code permission bypass is when a Bash command executes more -- or different -- code than what you approved in the permission dialog. The checker reads the command first, constructs what to show you in the dialog, then executes. If that construction step can be tricked into showing a sanitized version of a malicious command, the approval becomes meaningless -- a rubber stamp on something you never actually saw.
v2.1.223 is the fourth consecutive release in August 2026 to include permission-layer fixes. v2.1.221 patched a zsh regex conditional bypass that let hidden commands execute inside [[ ]] conditionals. v2.1.222 patched four security issues including git worktree path confusion (CVE-2026-55607), which allowed sandbox escape via symlink manipulation. The pattern suggests Anthropic is running systematic review across the permission and sandbox layers rather than reacting to individual reports.
Each bypass class in v2.1.223 is distinct in mechanism and attack surface. None required modifying Claude's model weights or jailbreaking the AI -- these were implementation vulnerabilities in the tool wrapper itself, which is both reassuring (model alignment still holds) and concerning (the tooling layer had real gaps).
Bypass 1: Command structure obfuscation
The first bypass let a crafted Bash command hide portions of itself from Claude Code's permission checker by exploiting how compound statements were parsed before the approval dialog was constructed. The hidden payload didn't appear in what you approved -- but it still executed. This affected any agentic session with Bash access, especially those driven by MCP tool outputs or external prompt injection.
An earlier variant of this class appeared in Claude Code v2.1.216: commands with more than 50 subcommands joined by &&, ||, or ; caused deny-rule enforcement to skip entirely. The v2.1.223 fix addresses a related structural variant -- per the official changelog: "fixed a Bash permission bypass where a crafted command could hide parts of itself from permission checks." The exact parser path it targeted wasn't disclosed, which is standard practice for security patches to prevent immediate re-exploitation.
The practical risk: in agentic sessions where Claude generates Bash commands autonomously, a malicious MCP server response or a prompt injection embedded in a file could craft a command that looked benign in your approval dialog while executing something else on your filesystem. With --dangerously-skip-permissions disabled (the default), this bypass could still produce a deceptive dialog rather than outright skipping it. You'd still click approve -- just on a sanitized fake.
Bypass 2: Invisible Unicode and tab padding
The second bypass used invisible Unicode characters -- zero-width spaces, directional markers, or non-printing characters in the U+200B to U+202F range -- or raw tab characters to pad a command so the approval dialog displayed a different string than what the shell executed. The display renderer would strip or collapse the invisible chars; the shell would not.
Security researcher Johann Rehberger had previously demonstrated that production agentic coding tools were vulnerable to invisible Unicode injection via skill files and MCP server metadata -- the payload hides in what looks like whitespace. Claude Code's GitHub issue tracker flagged this pattern in issue #29489: "Unicode confusable characters can bypass permission rule matching." The v2.1.223 changelog confirms the fix: "fixed permission prompts so commands padded with tabs or invisible Unicode can no longer hide part of the command from the approval dialog." The normalization now strips these characters before the display is constructed.
MCP servers are the primary attack vector here. Third-party MCP servers control the content of tool results returned to Claude. If a server returns a tool result containing invisible Unicode sequences, those characters could end up in a Bash command Claude generates. Before v2.1.223, that command would display cleanly in the approval dialog while executing a hidden payload in the shell. This is a supply-chain-adjacent risk: you don't have to install anything malicious -- you just have to connect to a server that returns adversarial content.
Bypass 3: Workflow dynamic import() sandbox escape
The third bypass is the most technically significant. Workflow scripts in Claude Code run inside a JavaScript sandbox that prevents them from directly calling Node.js APIs, accessing the filesystem, or spawning arbitrary processes. The fix closes a path where workflow scripts could use JavaScript's dynamic import() -- import(someComputedExpression) rather than static top-level import x from "y" -- to load modules from outside the sandbox boundary entirely.
Static imports are resolved at parse time and can be statically analyzed. Dynamic imports happen at runtime on computed values, making them fundamentally harder to sandbox without blocking the feature entirely. The vulnerability meant a workflow script could construct a path to a Node.js module outside the allowed scope -- think import("/home/user/.ssh/id_ed25519") or a path to a system module -- and execute arbitrary code at host-OS level, bypassing the sandbox guarantees completely.
The Workflow tool powers Claude Code's multi-agent orchestration: agent(), parallel(), pipeline(). The sandbox is what ensures workflow scripts can't directly mutate your filesystem or spawn processes outside the controlled agent() interface. CVE-2026-55607 (patched in v2.1.222) was a different sandbox escape via git worktree confusion. The dynamic import() escape is a separate path to the same outcome: code running at host level that should have been confined. The v2.1.223 fix either blocks dynamic import() in workflow context entirely or validates import targets against an allowlist -- the changelog doesn't specify which approach.
Get the AI Agent Briefing
One email per week. The best AI agent news, tutorials, and tools -- written by someone who actually builds with them.
Subscribe Free
The fourth fix: bypassPermissions policy enforcement gap
A fourth security fix in v2.1.223 closes a policy enforcement gap: agent definitions that explicitly set bypassPermissions mode were ignoring org-level policies that disable bypass. Enterprise organizations that configured managed settings to prevent any agent from skipping permission checks were not actually protected if an agent skill or slash command declared bypassPermissions in its own definition.
This matters specifically for teams using Claude Code in multi-developer environments with centralized policy management. The org-level bypass-permissions disable is meant to be a hard constraint -- no agent can run in bypass mode regardless of what its definition says. Before this fix, that constraint had a hole: agent definitions could declare bypassPermissions locally and override the org policy silently. With v2.1.223, the org policy enforcement is consistent across all code paths. If you manage Claude Code deployments for a team, this fix delivers on the security guarantee the managed settings UI implies.
What else shipped in v2.1.223
Beyond security, v2.1.223 adds owner wildcard entries for marketplace settings, surfaces the /teleport command in cloud sessions, and warns when a requested subagent model is restricted so the parent model silently runs instead. These are UX and configuration quality-of-life improvements, but each addresses real friction in multi-agent and enterprise deployments worth knowing about.
Owner wildcards for marketplace settings. Managed settings for strictKnownMarketplaces and blockedMarketplaces now support "owner/*" wildcard entries. Previously, allowing or blocking repos from a GitHub organization required listing each repo individually. With wildcards, you can write "myorg/*" once and cover all current and future repos under that org. For enterprise deployments where a trusted GitHub org is the primary source of Claude Code extensions, this is a practical reduction in maintenance overhead.
/teleport hint in cloud sessions. When running Claude Code in a cloud session at claude.ai/code, you'll now see an inline hint showing how to continue the session locally: claude --teleport <session-id>. Teleport pulls the full conversation history and branch state into your local terminal environment. Previously the command existed but wasn't surfaced in-session -- you had to know it was there. The v2.1.223 change makes it discoverable at the moment you'd want to use it: when you're already in a cloud session and want to move to your local machine with your full tool environment.
Model restriction warning. When a workflow agent, forked skill, slash command, or resumed background agent requests a specific subagent model and that model is restricted by org policy, Claude Code now shows a warning that the parent model is running instead of the requested one. Previously this silent fallback produced unexpected behavior in workflows built around specific model capabilities -- particularly for workflows that rely on extended thinking or specific context windows. The warning surfaces the substitution so you know what's actually running.
Bug fixes that matter for daily use
v2.1.223 resolves a regression present since v2.1.181: plugins enabled via the --settings CLI flag weren't loading at all. If you use custom settings files with plugin configurations and noticed missing functionality over the past several weeks, this is the fix. Long-running sessions now also correctly refresh feature flags when the OAuth token rotates -- previously, features could go stale in sessions running past the token expiry window, causing inconsistent behavior without an obvious error. Additional fixes address /ultrareview, the claude update and claude doctor commands, memory frontmatter value parsing, and telemetry double-counting.
How to update to v2.1.223
Update Claude Code to v2.1.223 by running npm update -g @anthropic-ai/claude-code or claude update from inside any Claude Code session. The release timestamp is 2026-08-06T00:52:37Z. Verify with claude --version -- it should show 2.1.223. If you run Claude Code in automated pipelines, crons, or CI environments, treat this as a priority update given the permission-layer and sandbox surface area covered in this release.
For teams using managed settings with the bypass-permissions disable policy, the v2.1.223 fix is particularly important -- before this release, that policy had an enforcement gap through agent definition overrides. If you've audited your agent definitions assuming the org policy was a hard constraint, that assumption is now correct for the first time.
FAQ
Is Claude Code v2.1.223 a required update?
It's not marked as mandatory, but it patches three distinct Bash permission-bypass classes plus a workflow sandbox escape and an org policy enforcement gap. If you run Claude Code in agentic mode with Bash access -- especially with MCP servers, third-party skill repos, or multi-agent workflows -- updating is the right call. Run claude update or npm update -g @anthropic-ai/claude-code to get v2.1.223.
Were these bypasses being actively exploited?
Anthropic has not disclosed active exploitation of the v2.1.223 vulnerabilities. The Unicode bypass class was flagged in GitHub issue #29489 as a publicly documented research finding. Security researcher Johann Rehberger demonstrated invisible Unicode injection in production agentic coding tools earlier in 2026. "Not publicly exploited" is not the same as "low risk" -- permission-layer bypasses in agentic coding tools are high-value targets for MCP-delivered prompt injection attacks.
What is the workflow sandbox in Claude Code?
Claude Code workflows run in a JavaScript sandbox that prevents workflow scripts from directly calling Node.js APIs, accessing the filesystem, or spawning arbitrary processes. All actions are supposed to route through the agent(), parallel(), and pipeline() primitives, which maintain their own permission checks. The dynamic import() fix in v2.1.223 closes a gap where runtime module loading could escape that sandbox boundary and execute code at host-OS level.
How often does Claude Code ship security fixes?
Based on August 2026 releases, roughly every other release includes a permission-layer or sandbox security fix. v2.1.216 patched compound statement and Unicode bypasses; v2.1.221 patched a zsh regex conditional bypass; v2.1.222 patched CVE-2026-55607 (git worktree sandbox escape) plus three other issues; v2.1.223 patches three more bypass classes. This cadence suggests active systematic audit rather than reactive patching -- which is the right approach given the surface area.
Get the AI Agent Briefing
One email per week. The best AI agent news, tutorials, and tools -- written by someone who actually builds with them.
Subscribe Free