INVERSEPROMPT: TURNING CLAUDE AGAINST ITSELF, ONE PROMPT AT A TIME (­­­CVE-2025-54794 & CVE-2025-54795)

Summary from the TLDR InfoSec newsletter

Elad Beber discovered two high-severity vulnerabilities in Anthropic's Claude Code during its Research Preview phase. CVE-2025-54794 (CVSS 7.7) allows path restrictions to be bypassed through naive prefix validation, enabling access to files outside the intended working directory. CVE-2025-54795 (CVSS 8.7) enables command injection via inadequate input sanitization, allowing arbitrary code execution without user confirmation by breaking out of allowed commands like “echo”. Anthropic has patched both vulnerabilities.