← Control library
S01 · Integration security
Untrusted input handling
A document, webpage, caller or tool result changes instructions and bypasses policy.
Requirement
Treat external content as data, preserve instruction priority and block content from granting itself authority.
Expected evidence
- Trust-boundary model and prompt policy.
- Injection test corpus with expected outcomes.
Proposed verification
- Run direct, indirect and encoded injection samples.
- Confirm blocked input cannot alter tools, recipients or approval state.
Limit
Adversarial techniques evolve; passing a fixed corpus is not permanent protection.