Vulnerability  ·  2026-08-15

Network-AI ClaudeHookBridge — command truncation and quote-stripping bypass allow dangerous Bash execution past Claude Code deny-list

VulnerabilityHigh impactGlobalCVE-2026-73614
NVD published CVE-2026-73614 and CVE-2026-73615 (both CVSS 8.8, High) on 2026-08-13, describing two distinct bypasses of Claude Code's Bash-command security policy enforcement in the ClaudeHookBridge plugin — one via string truncation before deny-pattern matching, one via quote-handling mismatch between the policy matcher and the executor.
These are agent-guardrail bypass vulnerabilities in a security layer specifically built to constrain what commands an AI coding agent can execute — defeating the very control meant to prevent dangerous command execution by Claude Code, a widely used AI coding agent.
ClaudeHookBridge truncates the target command string to 500 characters before evaluating denyPatterns, while Claude Code executes the full untruncated command; attackers position dangerous content past byte 500 in a Bash command field to bypass the deny-list. A related flaw (CVE-2026-73615, CVSS 8.8) has SandboxPolicy evaluate quoted commands while the executor strips quotes before execution, allowing quoted-command evasion of blocklist checks
Network-AI ClaudeHookBridge < 5.15.1
Upgrade to Network-AI ClaudeHookBridge 5.15.1 or later
GitHub Advisory GHSA-743h-jr5x-mpcrGitHub Advisory GHSA-9v4f-j8cv-fhxw
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →