intent-parser

Pass

Audited by Gen Agent Trust Hub on Jun 16, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: A vulnerability surface for indirect prompt injection exists in the intent classification workflow.\n
  • Ingestion points: Raw inbound Telegram messages are ingested as untrusted data in SKILL.md and passed to a classifier.\n
  • Boundary markers: The skill's algorithm describes passing user messages directly to an LLM classification prompt without the use of explicit boundary markers or instructions to disregard embedded commands.\n
  • Capability inventory: The output of this classification determines the routing to highly capable downstream tools mentioned in SKILL.md, including the Linear MCP for ticket creation, telegram-draft-for-review for communication, and dispatcher-core for executing multi-step flows.\n
  • Sanitization: Although regex-based PII filtering is performed, there are no specific sanitization or validation steps mentioned to prevent adversarial inputs from influencing the classification logic.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 16, 2026, 05:22 PM
Security Audit — agent-trust-hub — intent-parser