Runtime data lineage for AI agents

Security decisions based on where data came from.

AgentLineage tracks sensitive values through agent tools and generated Python code, then checks their provenance before network egress.

Get a demo Read the architecture

Technical preview · Python runtime · syscall-level fallback

agentlineage / decision

source    file:/run/secrets/api_key
label     SECRET
path      read -> strip -> encode -> json
sink      POST api.external.dev/upload

BLOCK SecretToNetwork
reason    secret-derived payload reached
          an unapproved network boundary

Enforcement model

source-awaretransformation-awaresink-enforcedexplainable

01 / Sources

Mark high-value inputs

File reads, environment values, tool outputs, and generated-code inputs enter the runtime with source labels.

02 / Propagation

Carry lineage forward

Supported string, byte, container, encoding, and serialization operations preserve origin and transformation history.

03 / Sinks

Decide before effect

Guarded HTTP, tool, subprocess, and code-execution boundaries evaluate the outgoing value before the effect runs.

04 / Evidence

Reconstruct the decision

Structured reports connect source, transformations, sink, policy, code hash, and event sequence.

Layered coverage

Precision in the runtime. Coverage at the system boundary.

Object-level dynamic taint tracking reduces process-wide false positives. Syscall-level process provenance remains the conservative safety net for subprocesses, native code, and unsupported paths.

Explore the architecture

01Framework adapterstool calls · model output · approvals

02Python taint runtimetracked values · generated code · guarded IO

03Policy + evidencedecision · explanation · JSONL events

04System provenanceprocess · file · pipe · socket