CVE-2026-54769

Name of the Vulnerable Software and Affected Versions Langroid versions prior to 0.65.2

Description A sandbox escape exists in the TableChatAgent and VectorStore capabilities that allows unauthenticated remote code execution (RCE) on the host system. The issue occurs when agents evaluate LLM-generated tool messages with full eval=True using Python's eval() function. While the framework attempts to sandbox execution by setting locals to an empty dictionary {}, it fails to scrub builtins from the globals dictionary mapping. Consequently, Python implicitly injects all built-ins, granting access to dangerous functions such as import ('os').system(). An attacker can exploit this by using prompt injection to force the TableChatAgent.pandas eval() function to execute a malicious payload.

Recommendations Update Langroid to version 0.65.2. As a temporary mitigation, avoid setting full eval=True when configuring TableChatAgentConfig until the update is applied.