Researchers discovered that most open-source AI agents use easily bypassed command filters (denylists) that fail to stop obfuscated destructive commands. Ten out of eleven tested open-source AI agents ...