Check Point Research has revealed critical vulnerabilities (CVE-2025-59536, CVE-2026-21852) in Anthropic’s AI coding tool, Claude Code. These flaws enable remote code execution and unauthorized access to API credentials, posing significant risks for developers. Simply cloning and opening untrusted projects can allow attackers to execute harmful shell commands and exfiltrate API keys without user action.
The analysis indicates a shift from traditional configuration files to active execution methods, where automation features like hooks blur the lines between setup and code execution. Key issues include hidden command execution, improper bypassing of user consent for integrations, and API key theft before trust verification.
As AI tools take on a more critical role in development workflows, security risks are escalating. Anthropic has addressed these vulnerabilities by enhancing user consent measures and blocking unauthorized actions. This highlights the need for evolving security frameworks to keep pace with AI advancements. Further details are available in the CPR blog.
Source link
