Agent prompt-injection and data-exfiltration tests

Create a repeatable security test suite proving that untrusted content cannot change an agent’s permissions or expose protected data.

Test instructions coming from:

• User messages.
• Knowledge documents.
• Web content.
• Tool and API responses.
• Encoded or obfuscated input.

Verify that this content cannot enable tools, change execution identity, bypass escalation, access unauthorised resources or extract protected information. Sensitive-data redaction and the product’s content-safety position should be separate tickets.