Guide-yakabatana mibvunzo · Zvakaoma Nhanho
The Dual-LLM Pattern Against Prompt Injection Quiz
Tests trust boundaries, limited extraction, validation, tool authorization, and the limits of dual-model defenses against prompt injection.
Mubvunzo 1 ye 8
Ndeupi modhi inobvumidzwa kuita yakatenderwa chishandiso chiitiko mupateni inotsanangurwa?
Tarisa mubvunzo wese8 mibvunzo
Mhinduro uye tsananguro zvinoonekwa paunenge uchitora bvunzo.
- Ndeupi modhi inobvumidzwa kuita yakatenderwa chishandiso chiitiko mupateni inotsanangurwa?
- The quarantined reader
- The privileged model
- Either model whenever confidence is high
- Neither; both only summarize
- Sei mirairo ine hutsinye yakadzikwa mupeji ichigona kukanganisa mumiriri wemhando imwe chete?
- Every page is executable code
- The context window automatically elevates page text
- The model may not reliably separate the page’s instructions from task instructions in its language context
- The page changes the model weights during inference
- Chii chinofanira kudzoka modhi yakavharirwa kuderedza njodzi paruoko?
- Only a narrow, schema-validated field needed for the task
- A tool call encoded as prose
- A broad narrative with suggested actions
- A raw copy of the page
- Mushure mekugamuchira mari yekubhadhara yakabviswa-modhi, chii chakakodzera cheki inotevera?
- Validate its type and permitted range in ordinary code
- Let it authorize its own payment
- Append it to the system prompt
- Trust it because a second model produced it
- Sei dhizaini yemhando mbiri isingavimbise kuti jekiseni rekukurumidza rinokundwa?
- It cannot use structured values
- The extractor or an indirect route may still pass hostile influence to the privileged model
- Two models always share weights
- Tool allowlists make injection worse
- Which operational cost comes with adding a quarantined model?
- It removes the need to evaluate the agent
- It eliminates all logging
- It adds model calls and integration work
- It prevents validating output
- How should CaMeL be described relative to the basic dual-LLM pattern?
- A classifier trained only on email
- A guarantee produced by a JSON schema
- A synonym for every two-model design
- Related research on explicit control/data flow and capabilities, with distinct system-level mechanisms
- Which diagram most helps review whether this architecture protects a tool-using workflow?
- A trust-boundary map of inputs, contexts, permissions, state, and handoffs
- A list of prompt templates without data paths
- A screenshot of the user interface
- A chart of model parameter counts only