What attackers do. This item maps to MITRE ATLAS, the knowledge base for attacks on AI systems, where the technique covers crafted prompts meant to push a large language model outside its intended behaviour and safety controls. Attackers use role play, encoding tricks or layered instructions so the model produces content or takes actions it should refuse.
How to defend. Do not rely on model guardrails alone: add independent input and output filtering, restrict what tools and data the model can use, and apply least privilege to any actions it can take. Keep logs of prompts and responses, red team applications with known bypass patterns, and update protections as new methods appear.
Signs to look for. Look for prompts with unusual encodings, long role play setups, repeated rephrasing after refusals, and outputs that violate policy. A common gap is monitoring only for harmful output while ignoring repeated probing attempts.