Trust & security

Every regulatory date here has a source and a check date.

The dates come from one register that records the provision behind each date, its EUR-Lex source and the day it was last read against that source. Latest check: 25 Oct 2026. See the timeline.

How your data is handled

Four guarantees, by architecture, not by promise.

These are not policy commitments we ask you to trust. They are properties of how the tools are built, verifiable in the open-source code before you install anything.

  • Runs locally

    Executes on your infrastructure and reads your code. You control where it runs and what it touches.

  • Zero telemetry

    No phone-home and no usage tracking. The tools never report back to us.

  • No data leaves

    Your code and the artefacts the tools emit never reach us. There is no server to send them to.

  • Apache 2.0

    Full grant, patent included. Read every line before you trust it, then run it under your own terms.

Security posture

How the tools are built to be safe to run.

Each control below carries a proof-tag pointing at the mechanism that enforces it, in CI or in the CLI, not at a paragraph of reassurance.

  • Network use, tool by tool

    LitmusAI and RiskForge make no outbound calls by default, enforced by a CI socket gate. License Compliance Checker looks up licence data online unless you pass --offline. RAG Benchmarking judge metrics call the judge model you configure.

    pytest-socket · CI-gated

  • No accounts, keys only for model calls

    No tool needs an AiExponent account, a sign-in or a licence server. LitmusAI, RiskForge and License Compliance Checker need no API key (LCC only for its optional LLM analysis). RAG Benchmarking judge metrics need GEMINI_API_KEY or OPENAI_API_KEY, and Agentic Document Analyser needs FIREWORKS_API_KEY.

    pip install · run

  • Tamper-evident audit log (RiskForge)

    Every change made through RiskForge is recorded in a hash-chained log. riskforge verify flags hand edits and cut or deleted log entries, but not a log cut back or rewritten with its head file edited to match. A receipt kept outside the project catches that for every entry up to the receipt.

    SHA-256 chained

  • Auditable supply chain

    Public source, pinned dependencies and CycloneDX SBOM output let you vet exactly what you run before it touches your systems.

    SBOM · pinned deps

Standards alignment

We align with the frameworks. We don't claim certification.

Tool outputs map to recognised control frameworks so your evidence is portable. Alignment is not the same as a certification, and we won't imply it is.

  • NIST AI RMF

    RiskForge dimensions map to GOVERN / MAP / MEASURE / MANAGE, so a single assessment serves both the EU AI Act and the US framework.

    Aligns, does not certify

  • ISO/IEC 42001

    Outputs map to Annex A controls to support an AI management system, increasingly a procurement gate for enterprise buyers.

    Aligns, does not certify

Responsible disclosure

Found something? Tell us privately first.

Report a security issue to security@aiexponent.com. We acknowledge reports, coordinate a fix, and credit reporters who want it. Machine-readable contact details are published at /.well-known/security.txt. For non-security bugs, use the public GitHub issues.

Disclosure contacts

See also the legal notices for the warranty disclaimer.

Related

Last reviewed: 2026-07-14