Anthropic’s Claude Desktop has been found to contain critical remote code execution (RCE) vulnerabilities in its official extensions for Chrome, iMessage, and Apple Notes, primarily due to unsanitized command injections. These vulnerabilities, which carry a high severity score represented by a CVSS of 8.9, could allow attackers to execute arbitrary code on users’ systems through innocuous interactions with the AI assistant. KOI Security discovered these flaws, which highlights the risks present in emerging AI ecosystems where extensions interact directly with powerful language models and local systems without adequate safeguards. Unlike traditional browser add-ons that operate with sandboxing, these extensions run unsandboxed on user machines, granting them significant access to files and commands.
Each extension processes user inputs, enabling potential exploitation through malicious command injections. A specific exploit involves an attacker crafting inputs that escape string context and inject arbitrary AppleScript, which can lead to elevated shell command executions. Moreover, the real threat emerges when attackers manipulate web content that Claude Desktop interacts with, like tailored search results, to seamlessly execute unauthorized commands. The vulnerabilities exposed by these popular extensions stress the need for stricter security practices, especially in an environment where independent developers rapidly deploy AI-related code. Anthropic acted promptly to resolve these issues, but heightened vigilance in the use of such powerful tools remains essential for preventing security breaches.
👉 Pročitaj original: Cyber Security News
