AI safety is not limited to frontier research. Everyday teams need practical controls for incorrect output, data exposure, unsafe actions, manipulation, and overreliance.
Separate data from instructions
Web pages, emails, uploaded documents, and tool responses can contain hostile instructions. Systems should treat that content as evidence to analyze, not authority to change settings, reveal secrets, or contact people.
Limit authority
Give each workflow only the credentials and tools it needs. Require approval for payments, publishing, account changes, and destructive actions. Apply rate limits, budget limits, and destination allowlists where appropriate.
Test realistic failures
Evaluate more than ideal prompts. Include ambiguous requests, missing data, prompt injection, unavailable tools, and malicious files. Record why a workflow stopped or escalated so teams can improve it.
For link workflows, inspect unfamiliar destinations with the link safety checker. No automated signal guarantees safety, but layered checks reduce avoidable risk.