.svg)
What is Deepset Injection Attack?
A Deepset Injection Attack is a prompt-injection evaluation that stresses an AI system with crafted instructions--often in user text or retrieved documents--to override system intent, exfiltrate context, or force unsafe tool use.
What this looks like in production
In RAG or agent stacks, a document that says ignore previous rules can rank highly and steer the model. Chat-only apps see the same pattern via direct user messages.
What teams usually do about it
- Treat retrieved and user text as untrusted instruction channels.
- Separate system policy from untrusted content.
- Include injection suites in continuous testing.
Related Giskard articles
Test injection resilience with Giskard
Map Deepset-style injection cases to Giskard prompt-injection and RAG probes. Start with the dynamic red-teaming scanner.
Further reading
Authority reference: OWASP LLM01 Prompt Injection.
