AI agents do not need Hollywood hacking if your keys are lying around.
India's CERT-In says AI agents from multiple providers reached live systems at Indian organizations during research and testing, according to TechCrunch. The entry points were boring and dangerous: exposed credentials in app configs, public repos, connection settings, and cloud roles with too much access. In some cases, agents read personal data and confidential business records.
My angle for small business automation is simple: before you connect an AI agent to leads, invoices, documents, warehouse data, or support tickets, check the blast radius. A Claude-based pipeline, custom OCR flow, or n8n scenario should have a separate service account, minimum permissions, secret storage, logging, and a kill switch. Otherwise you are not buying automation - you are giving a very fast intern the master key.
I'd say this honestly: agents are useful when the process is narrow and the access is narrow. Invoice parsing can read one mailbox and write to one table. A document agent can see one folder, not the whole drive. This is the difference between hours automation gives back and a security incident that does not pay for itself.
Why it matters: AI agent security is now an operations task for small businesses, not only a Big Tech problem. Before any prototype goes live, permissions should be designed as part of the workflow, not cleaned up after.
Source: TechCrunch