Controlled Research Prototype — Synthetic Environment

Stratyx Interlock · Research prototype

AI agents can request. Interlock decides what executes.

Interlock is a proposed independent authorization and containment layer that evaluates each consequential agent action before it executes — across delegation handoffs and the full sequence of prior actions.

Enter the Interlock Lab Request a Research Briefing

The problem

Valid identity is not valid intent.

Identity and authentication establish who an agent is and which resources it can reach. They do not establish whether a specific action is still consistent with what a human actually delegated — especially after an agent hands work to sub-agents, reads untrusted content, or accumulates context over many steps.

A credential can remain valid while the action it authorizes has drifted far from the original task. Interlock is researching how to evaluate that gap at the moment of execution, deterministically, and fail safely when evidence is missing.

Five decision outcomes

Allow

The action matches delegated intent, scope, sequence, and conditions. It may execute.

Deny

A hard policy check failed. The action does not execute.

Require Approval

The action may be legitimate but needs fresh, independent human approval bound to this exact request.

Contain

The workflow is paused and its reach restricted before a consequential step executes.

Revoke

The delegated authority behind the request is withdrawn for the agent and its descendants.

Evaluated in order

  1. 01

    Identity and attestation

  2. 02

    Delegation-chain validity

  3. 03

    Task and intent consistency

  4. 04

    Tool and resource scope

  5. 05

    Parameter boundaries

  6. 06

    Sequence and cumulative risk

  7. 07

    Approval provenance

  8. 08

    Operating conditions

  9. 09

    Reversibility

  10. 10

    Policy-engine health

A hard policy failure always overrides a favorable score. Outages, timeouts, and missing evidence never default to Allow.

Controlled synthetic scenarios

  • S1

    Normal authorized action

  • S2

    Direct prompt-injection attempt

  • S3

    Forged human approval

  • S4

    Harmful multi-step sequence

  • S5

    Excessive agent delegation

  • S6

    Multi-agent task drift

  • S7

    Authorization-system outage

Current status

Research prototype in development

Stratyx is validating its threat model and building a controlled synthetic environment to test whether the enforcement model holds up under adversarial scenarios. We are seeking expert feedback and design-partner conversations.

Prototype limitations

  • All agents, people, tools, resources, and actions are fictional and seeded for demonstration.
  • No connection to real infrastructure, credentials, payment systems, or operational technology.
  • Containment and revocation are simulated inside the application only.
  • Decisions are deterministic rules evaluated server-side; no language model makes the final decision.
  • Audit hashes demonstrate integrity concepts and are not production-grade non-repudiation.
  • The prototype has not been independently validated and is not production-ready.
Enter the Interlock Lab