Test, harden and evidence AI assistants before release.
HexaSec AI Assurance Gate provides repeatable security regression testing for LLM and RAG-enabled assistants, producing deterministic gate decisions and audit-ready evidence.
AI assistants are connecting to everything — with very little evidence behind them.
Before release, organisations need evidence that their assistants will not leak sensitive data, follow poisoned retrieval content, ignore policy or trigger unsafe tool actions.
AAG provides that evidence on every change — not just at launch.
Direct and indirect injection through inputs, documents, search results and tool outputs.
Sensitive content surfacing in responses, citations, retrieval traces or tool calls.
Adversarial content embedded into corpora or hosted documents to alter behaviour.
Assistants calling tools in ways the policy forbids — quietly, at scale.
Eight things, deterministically.
Scenario packs
Structured, versioned packs of LLM and RAG security scenarios you can extend.
Deterministic detectors
Pattern, structural and policy checks — not opaque judge-model scores.
Policy-as-code
Explicit rules for what assistants may say, retrieve, cite or call.
Tool-action validation
Catch unsafe, out-of-policy or unintended tool calls before they execute.
Retrieval trace analysis
Inspect what was retrieved, ranked and used — not just what was said.
Evidence pack generation
Signed bundles for assurance, procurement and change control.
Gate decision output
GO · CONDITIONAL · NO-GO — reproducible across runs.
Re-test on change
Re-run the gate on every model, prompt, policy, tool or corpus change.
A repeatable assurance flow.
From adapter to evidence bundle — deterministic at every step.
Connect
Attach to the assistant through a local adapter.
Run packs
Execute structured security scenario packs.
Capture
Record responses, retrieval events and tool attempts.
Check
Apply deterministic detectors and policy rules.
Decide
Produce a reproducible gate decision.
Evidence
Generate a signed evidence bundle.
Re-test
Re-run on every relevant change.
Three states. No grey-zone judgement.
Every run produces a single, reproducible decision tied to the bundle of evidence behind it.
All scenarios passed and all policy checks were satisfied. The assistant meets the configured assurance bar.
Acceptable risk profile with named exceptions. Owner sign-off required against tracked conditions.
Material failures, unsafe behaviours or policy violations were detected. Release is blocked pending remediation.
A defensible surface, not a buzzword list.
AAG ships with structured packs covering the failure modes that matter for LLM and RAG-enabled assistants in regulated environments.
What a run leaves behind.
Every gate decision is anchored to a signed evidence bundle that an external reviewer can inspect — without needing access to the live system.
DIGEST · sha256:7a4f…9e02
What AAG is — and what it isn't.
Honest scoping makes assurance work. AAG is a focused tool; some adjacent jobs belong elsewhere.
- An assurance gate for LLM & RAG assistants
- A security regression test harness
- A deterministic evidence generator
- A local-first testing tool
- A pre-release and post-change control mechanism
- A chatbot or general LLM platform
- A SOC or SIEM platform
- A GRC platform
- A replacement for human approval
- A live offensive red-team toolkit
Run AAG against your assistant.
Pilots are scoped, time-boxed and run against your environment. We share the pilot pack, one-pager and technical documents on request.