sms
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFEDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: Untrusted SMS message content is ingested from
inbox.jsonl(populated bysms_listen.py) and interpolated into prompts for theclaudeCLI insms_respond.py. An attacker could send crafted SMS messages to influence the agent's persona (soul state) or poison user profiles stored insms_memory.db. Whilesms_respond.pyuses backtick delimiters and explicit instructions to the LLM to ignore markup in the message, the attack surface for indirect injection is present.\n- [COMMAND_EXECUTION]: The skill frequently executes shell commands viasubprocess.run.sms_respond.pycalls theclaudeCLI to generate auto-replies, andsms_process_reply.pyspawnssms_send.pyas a subprocess. These are intended behaviors but represent high-privilege operations within the skill's scope.\n- [CREDENTIALS_UNSAFE]: The shared utility script_sms_utils.pyreads sensitive credentials from~/.claude.jsonand~/.config/env/secrets.env. This allows the skill to automatically retrieve API keys for Twilio and Telnyx. Accessing these files is a form of local credential harvesting required for operation.\n- [DYNAMIC_EXECUTION]: The listener and responder daemons (sms_listen.pyandsms_respond.py) useos.fork()and related syscalls to run as background processes, a technique used for persistence and standalone operation on the host system.
Audit Metadata