ENGINEERING FIELD NOTE / PromptShield
Deterministic prompt-risk scanning
A small rule set flags instruction overrides, jailbreak markers, and exfiltration intent with evidence excerpts and remediation guidance — an auditable first-pass filter, honestly labelled heuristic.
Return evidence rather than a bare verdict
PromptShield scans prompt text with deterministic rules and returns severity, findings, evidence excerpts, and remediation guidance. A reviewer can trace a result to the matching text instead of relying on an unexplained safe or unsafe flag.
Check benign language as well as attacks
The rules cover instruction overrides, role hijacking, jailbreak markers, exfiltration, and unsafe tool-use instructions. Benign security discussions can contain the same vocabulary. A regression test ensures ordinary API-key rotation policy language is not treated as data exfiltration.
Use the scanner as one layer
Regex rules are transparent and easy to test, but they can miss paraphrases and flag legitimate text. The scanner is a first-pass filter. Tool authorization, output validation, sandboxing, and human approval need their own boundaries.