Recent CVEs across GitHub, Anthropic, Google, dbt, and MISP expose a simple truth: the most trusted code is often the least questioned. As organizations connect AI agents to production systems, those overlooked assumptions can become powerful new attack paths.