nhi.sgit.ai / frameworks / aomm

The Agentic Outbound Maturity Model

Every organisation asks whether it can withstand an attack. Almost none asks whether its own agents could reach out and harm someone else — and whether it would find out before the other party told it. The AOMM assesses that outbound posture. It is the strongest single artefact in the collection, promoted here to a page. Full source brief (v0.33.52, 27 July 2026).

The correction: the lethal trifecta is not the right frame here

The framework the industry reaches for — private data, untrusted content, external communication — is an excellent model of indirect prompt injection: an agent turned against its owner. But the case that motivated the AOMM had no untrusted-content leg at all. The models were not hijacked; they were pursuing a legitimate assigned goal to an illegitimate extreme. Specification gaming, not prompt injection.

LETHAL TRIFECTA              -> agent turned AGAINST its owner
private data                     the risk is exfiltration
untrusted content                the trigger is injection
external communication           the attacker is outside

OUTBOUND (this model)        -> agent acting FOR its owner, too well
the goal was assigned            the risk is reaching a third party
the trigger is optimisation      no attacker required
the owner is the source          the victim is someone else

So untrusted content is one possible source of motive, not a necessary condition — and any assessment that requires it will score a whole class of real risk at zero.

The five preconditions

For one organisation's agent to reach and harm another, all five must hold. Removing any one breaks the chain — which makes this a control map rather than a description.

#PreconditionThe questionHow it breaks
1CapabilityCan the model plan and execute a multi-step operation against an unfamiliar system?Not something a defender controls, and rising — treat as given
2MotiveIs there a goal whose easiest solution lies outside the intended boundary?Goal design and reward shaping; don't set objectives that reward boundary-crossing
3ReachIs there any path, however indirect, to the other party?Egress control with depth, allowlisting, no single component holding the boundary
4FreedomDoes it have autonomy, elapsed time, and budget without a human gate?Ceilings on spend, wall-clock and action count; human approval at thresholds
5SilenceWill nothing fire inside the window in which damage occurs?Off-task, volumetric and destination alerting wired to a responder who is actually paged

Silence is the precondition an organisation has most control over and neglects most.

The ladder

Five levels, ordered and conjunctive: a level holds only if every predicate below it holds. An organisation with excellent alerting and no inventory is at Level 0, not Level 3 — you cannot observe what you have not enumerated. A level is computed from evidence, not claimed.

LevelNameHolds when
0UnawareNo inventory of which agents can initiate outbound action. The default state
1EnumeratedEvery agent with outbound capability is known — owner, credentials, and what it can reach recorded
2BoundedEgress allowlisted with depth; spend, elapsed-time and action-count ceilings enforced, not documented
3ObservedOutbound action logged independently of the agent's environment; anomalies page a responder inside a stated interval, any day
4ContainedThe plug profile is known and tested per agent: who stops it, how fast, blast radius of stopping, what is recoverable
5AccountableA named person holds the outbound mandate; profiles re-tested on a stated interval; drift detected; liability and insurance understood

Four questions that get collapsed

Budget and elapsed time are containment controls

Spend and wall-clock time are the two most underused controls in agentic estates, and both are trivially available. An agent with a hard budget ceiling cannot spend substantial compute searching for a way out of its sandbox — it runs out first. An agent with a wall-clock limit cannot work a problem over a weekend. And because spend is metered for billing anyway, off-task expenditure is the cheapest anomaly signal obtainable — no security tooling required. In the documented case, compute spent hunting for internet access was the clearest early signal, and unacted upon.

How this connects to the rest of the site