Error Message Mining

LLM API error responses are generated by multiple layers — the model itself, the inference framework, the API gateway, the safety classifier, and the orchestration layer. The design assumption is that error messages provide enough information for debugging without leaking sensitive internal details. The gap: each layer produces error messages independently, and their composition reveals implementation details.

• Monitor for systematic error-probing patterns: many 4xx/5xx responses from a single source
• Alert on requests designed to trigger errors (null models, extreme parameters, malformed JSON)
• Detect sequential boundary-probing patterns (incrementing token lengths, varying payload structures)
• Log and flag path traversal patterns in model name fields

Error message mining is the reconnaissance technique that feeds all other T5 techniques. Framework identification feeds T5-AT-006 (API Endpoint Abuse with known endpoint patterns).