interactive-bot
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill implements a two-tier command system where free-text messages are routed through an AI model (HAIKU_MODEL) with access to Gmail MCP tools (list, search, read, draft). This creates a vulnerability surface where an attacker sending an email or a Telegram message could inject malicious instructions into the bot's processing flow.
- Ingestion points: Untrusted free-text messages from Telegram are processed by
scripts/lib/agent-router.tsand passed to an LLM. - Boundary markers: The skill mentions "Anti-contamination" detection, but the implementation details are not visible in the provided files.
- Capability inventory: The bot has access to powerful Gmail tools (list, search, read, draft) and the agent reading the skill has access to
Bash,Grep, andGlob. - Sanitization: The skill mentions HTML escaping in
scripts/lib/telegram-format.ts, but this is primarily for output rendering rather than input sanitization. - [COMMAND_EXECUTION]: The skill uses the
Bashtool in its "Mandatory Preflight" section to check for the existence of scripts and verify specific exports in the plugin directory. While localized, this demonstrates the agent's ability to execute shell commands. - [PROMPT_INJECTION]: The "Self-Evolving Skill" section and the "Post-Execution Reflection" block instruct the agent to modify the
SKILL.mdfile immediately if instructions are found to be wrong. This self-modification loop could be exploited to persist malicious instructions or bypass original constraints if the agent is tricked during an interaction.
Audit Metadata