Skip to content

Pipeline

The pipeline is TapPass’s core enforcement mechanism. Every AI agent request passes through a configurable sequence of 44 security steps in three phases. A block at any step stops the request immediately (fail-closed).

PhaseStepsPurpose
Before the LLMInput validation, PII, injection, exfiltration, memory poisonScan and block before LLM sees the request
The CallLLM/tool execution, permissions, constraints, approval gateExecute with guardrails
After the LLMOutput scan, DLP, taint check, shell bleed, cost trackingScan the response before returning to agent

All regex and Presidio patterns run once on input text. Detection steps read from the cached ScanResult. Latency: ~250ms (down from ~2500ms with independent scanning).

Deploy new policies in shadow mode: the pipeline runs all steps and logs what would be blocked, but never actually blocks. Tune thresholds before enforcement.