Your AI controls are approved. Do they work?
Lantern tests whether the controls on your AI agents and AI-driven decisions actually do what your organisation approved, and gives risk, internal audit and regulators signed evidence they can check themselves.
Why now
Australia's prudential regulator wrote to every regulated entity on 30 April 2026. Its message, paraphrased:
Assurance isn't keeping pace with AI. Point-in-time, sample-based testing doesn't suit systems that learn, adapt and act on their own.APRA letter to industry on AI, 30 April 2026 (paraphrased)
Second-line risk and internal audit are expected to have the capability and tooling to independently assess AI systems, including agentic workflows.APRA letter to industry on AI (paraphrased)
Boards rely too heavily on vendor presentations, and access controls haven't caught up with AI agents acting as non-human users.APRA letter to industry on AI (paraphrased)
Read the letter in full at apra.gov.au. Your platforms enforce controls; Lantern is the independent check that those controls match what was approved.
What Lantern tests
Each test starts from a control objective your organisation approved, in an interpretation a named reviewer confirms. Only confirmed objectives are tested.
Policy enforced as approved
Boundary cases run against the rule your platform actually enforces. A refund limit of $10,000 should hold at $10,001, not $10,500.
No unguarded paths
The agent's identity can reach its tools only through the control point, not around it.
Limits that can't be split
Per-action limits that an agent can defeat by splitting one payment into three, shown with evidence from the activity log.
Separation of duties
An agent can't approve the actions that were held for human review.
Concentration and fallback
Critical agents that depend on one provider have a tested fallback.
Unapproved AI
Staff use of unapproved AI tools is prevented, not just detected after the fact.
How an assurance sprint works
Eight weeks, fixed fee, run inside your environment. Nothing needs to leave it.
Scope and export
Pick two or three agents. Your team exports their policies, enforced rules, permissions and activity log to files.
Confirm objectives
A named reviewer in risk or audit confirms the interpretation of each control objective before anything is tested.
Test
Lantern runs on your machine, with no network connection, and tests every confirmed objective.
Evidence and findings
A signed evidence pack and findings register your auditors can verify offline, and a walk-through with your risk and audit teams.
Independent by design
- Lantern never writes, changes or enforces a control, so it never marks its own homework.
- Every connection is read-only. Lantern never stores prompts, messages or model inputs and outputs, only metadata and signals.
- Every finding carries the evidence and test cases needed to reproduce it.
- Evidence packs are signed (ECDSA P-256 and ML-DSA-65), so anyone with Lantern's public key can check a pack hasn't changed since it was issued.
- Lantern provides evidence. It doesn't issue audit opinions or certifications: your auditors decide.
Talk to us
We're running the first assurance sprints with Australian banks, insurers and super funds. If your team is working out how to test AI controls after APRA's letter, we'd like to hear how you approach it today.