Australian not-for-profit · ACN 699 651 771Open research · Open software · Public benefitPublic record

Learn / AI safety

How an autonomous agent stays within bounds

Test a proposed action against authority, resource ceilings, and containment.

DSEMA research architecture · Reviewed 3 October 2026

Explore the mechanism

Try an agent actionDSEMA design • illustrative gates

Being able to pay is only one condition.

  1. 01

    Authority

    Pass
  2. 02

    Resource budget

    Pass
  3. 03

    Containment

    Pass
May proceed within scope

All three illustrative gates pass. The analysis may run within its approved scope. Passing these gates does not prove the output is correct or the system aligned.

Predetermined scenarios illustrate necessary checks; this does not execute an agent or verify enforcement. Source: DSEMA safety research.

01 / Explanation

Separate useful from permitted

An action can be profitable or technically successful while still exceeding its authority. DSEMA investigates checks before execution, restricted execution environments, external oversight, and controlled amendments. These are design requirements to evaluate, not a demonstrated general solution to AI alignment.

02 / Explanation

A budget is one gate

A compute ceiling constrains one resource channel. It does not alone stop data exfiltration, manipulation, collusion, or execution on infrastructure outside the controlled system. Enforcement must bind scheduling, credentials, network access, model loading, and the actual execution environment.

03 / Explanation

Measure failure under adversarial conditions

Test attempts to exceed a budget, move outside containment, spoof authorisation, compromise overseer keys, and game evaluations. Record whether execution was prevented, the evidence retained, and whether legitimate tasks were wrongly blocked. Bosun provides a narrower engineering track; its success would not establish all DSEMA controls.

Check your understanding

An agent has enough money for an unauthorised action. Should that action pass?

Next: From a useful task to paid computation →All guides