When robust input controls are not practical on a large language model (LLM) to prevent...

When preventive input hardening isn’t feasible for LLMs, AAISM prescribes compensating detective and corrective controls—notably human review and annotation of outputs prior to downstream action—to reduce harm from prompt injection. Output-side review gates prevent untrusted instructions from propagating, enable rapid suppression/feedback loops, and provide labeled examples for subsequent model hardening. IAM (B) is necessary but does not mitigate injection in content; reviewing inputs (C) is less effective than auditing what the model is about to act on; fine-tuning for validation (D) is helpful long-term but is not an immediate compensating control when robust input validation is impractical.