Keep evidence inside the boundary
Run detection and scoring in-process, with policy, trace and report digests joined at the decision point.
Security & privacy architecture
Privacy testing that keeps evidence inside your boundary.
AgentLeak is local-first, deterministic and redacted by default. Use the open-source analyzer offline, self-host the full platform, or send synthetic traces to the hosted service.
The security control loop
Build privacy controls around evidence.
Read regressions before release
Compare exposure over time and see when a new run crosses the project’s defined privacy threshold.
See every boundary crossing
Follow a redacted value from its source to the disclosure event without opening another project’s evidence.
Local by default
The core detector and AgentRisk scorer run in-process with regex, dictionaries, entropy and optional Presidio. No account, network call or hosted model is required.
- No telemetry or phone-home
- Raw matches redacted in reports
- Synthetic scenarios for safe validation
Trace-linked evidence
Every finding keeps its channel, event and severity context. Reviewers can see where data entered, which agent handled it and where disclosure occurred.
- Eight normalized execution channels
- Leak provenance and topology
- Stable report digests for CI
Bounded automation
Projects define a vault, detectors and policy. Autonomous agents receive scoped keys and machine-readable remediation without access to another project’s evidence.
- Per-project credentials
- Bring your own model key
- Explicit quotas and rate limits