Vulnerability  ·  2026-11-15

Claude Code v2.1.325 (2026-11-13) fixes Bash permission-rule / approval-gate bypass and sandbox hardening gaps

VulnerabilityMedium impactGlobal
The Claude Code v2.1.325 release (2026-11-13) shipped a cluster of security fixes in the coding agent's command-permission and sandbox boundaries: Bash allow rules could permit a program different from the one written in the rule, some compound commands executed without confirmation, deny/ask rules were missed on some Bash commands, relative-path deny/ask rules no longer applied after a cd, and a security issue in decoding zstd-compressed responses was patched (details undisclosed in the changelog). It also fixed a data-flow issue where token counting could send the conversation to Amazon Bedrock even when the organization's allowedProviders setting did not allow it.
Claude Code is a high-trust AI coding agent whose command-approval gate is the main defense between a prompt-injected or malicious repository and the operator's machine/cloud credentials. A bypass in how Bash allow/ask/deny rules are matched (authorizing the wrong program, letting compound commands through unconfirmed) directly undermines that gate, and the zstd/decoding and Bedrock data-flow fixes point to additional trust-boundary defects in the agent's response handling.
Bash permission-rule and sandbox approvals in the agent's command-approval gate were not reliably enforced: allow rules could authorize a different program than the one written in the rule, some compound commands ran without confirmation, and deny/ask rules were not applied to some Bash commands (and relative-path deny/ask rules stopped matching after a directory change). An attacker who can influence repo contents, hooks, or agent tool calls could get commands executed that the operator's policy was meant to deny. The release also fixes a security issue in zstd-compressed-response decoding and a data leak where token counting sent the conversation to Amazon Bedrock even when the managed setting allowedProviders forbade Bedrock.
Anthropic Claude Code before v2.1.325 (released 2026-11-13); also relevant to the bundled Claude Agent SDK / settings behavior for Bash and sandbox permissions.
Update to Claude Code v2.1.325 (2026-11-13) or later. Review permission rules after upgrade: sandbox Bash now follows the normal permission flow by default — set sandbox.autoAllowBashIfSandboxed: true only if you intended automatic sandboxed-Bash approval, and re-check allow/ask/deny rules for program-path mismatches.
anthropics/claude-code releases — v2.1.325Key Updates in Claude Code v2.1.324–v2.1.325 (2026-11-13)
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →