Identify how an uploaded image can become untrusted instruction content in a multimodal AI workflow.
The evaluator can read the image and update the score. Where is the actual security issue? Uploaded screenshot content Correct. The screenshot is untrusted content. In a multimodal workflow, image content can carry instructions the model may parse. Fixed QA rubric Not the main issue. A fixed rubric is a guardrail; the problem is letting image content compete with that rubric. Tester at the laptop Not the issue by itself. The human is using the tool normally; the risk is the untrusted media entering the model's instruction context. Score write action Important but incomplete. Tool access increases impact, but the…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in