How the attack works
An attacker sends a natural-language instruction to an agent that has database or tool access. The instruction asks for a broad, unqualified action -- delete everything, show every record, insert data into 'the most appropriate table' -- without naming any specific table, schema, or record. The agent translates this vague intent into an actual destructive or bulk-disclosure database call. No SQL injection syntax is used at all; the attack relies entirely on the agent's own tool-calling ability to turn plain words into a broad, unscoped action.
Why it matters
An organization can lose entire datasets to deletion or have all user records exposed in one action, triggered by a single unscoped sentence rather than a technical exploit.
What you can do
- Require agents to reject or ask for confirmation on database instructions that lack a specific, named table, schema, or record scope.
- Enforce least-privilege database credentials for agents so no single tool call can affect all tables or all users.
- Log and review agent-issued database operations, flagging ones with broad qualifiers like 'all', 'every', or 'most appropriate'.
- Add scope-validation logic between the agent and the database layer so destructive or bulk-read operations require an explicit allowlisted target.
Known benign look-alikes
- A developer or DBA giving a specific, scoped instruction naming an actual table/database ('drop the temp_orders table in staging')
- Documentation or tutorials explaining general database concepts without directing an agent's tool to act
- A legitimate data-export request scoped to a named report or a single user's own account, not 'all users'
- A request that says 'all tables/databases' but explicitly scopes the whole request to a non-production environment -- 'Truncate all tables in the test database before running the integration suite', 'Please drop all tables in the dev schema and recreate them for the new test run' -- must not fire, as long as the request does not ALSO reference production/live data anywhere