Prompt Inspection

Roadmap

Inbound prompts are scanned for injection patterns. Honest framing everywhere: this detects and contains — it cannot fully prevent.

Inactive

Input & Output Integrity

medium

P12

Satisfies

OWASP LLM

OWASP ASI

EU AI Act

Underlying control— show

content: role in (user, system) and label contains "injection"edit as custom control →

Injection heuristics

Detectors toggle individually; hit counts are the last 24h.

Instruction override

medium

2 hits"ignore previous instructions" family

Role confusion

medium

0 hits

System-prompt probe

high

1 hit

Encoded payload

medium

0 hitsbase64/hex smuggling — experimental

Scoped carve-outs with a reason and an expiry — accepted risk, on the record.

Scope	Value	Reason	Expires
No exceptions — the policy applies everywhere.

Remediation shown on findings: Review the flagged prompt; quarantine the source if hostile.