top of page

Agentic AI Control Catalogue 3.0

Version 3.0.0 · published 26 October 2026 · 143 controls · 10 families

143 controls for governing AI agents in production. Each control states a mandatory requirement, the failure it prevents, and whether it is technical or procedural.

What is published here

NPW publishes the list of controls openly. Verification methods, applicability rules and mappings to AI governance frameworks are NPW proprietary information. They are available through NPW consulting engagements.

​

Apply this list to your own agent. The Agent Governance Platform selects the controls that apply to one agent and maps them to the framework you choose.

Control ID
Family
Control
What it addresses
Control type
AI01
Agent Identity & Authorisation
Each agent must have its own identity, which no human user and no other agent holds.
Actions cannot be attributed to a single agent because identities are shared.
Technical
AI02
Agent Identity & Authorisation
Credentials issued to an agent identity must be bound to that identity and must be rejected when presented by any other principal.
A compromised secret gives access to several agents or users.
Technical
AI03
Agent Identity & Authorisation
The platform must issue agent credentials at task start, and the credentials must expire no later than the end of the task.
Static credentials remain valid after compromise and after the task has ended.
Technical
AI04
Agent Identity & Authorisation
Each agent must hold no standing permissions beyond those required by its declared task.
Agents hold broad permissions because precise scoping was harder than granting them.
Technical
AI05
Agent Identity & Authorisation
Where an agent's credentials cannot be issued at runtime, they must be rotated on a defined cycle and immediately on change of agent owner, decommissioning or suspected compromise.
Long-lived static keys outlive the agent, its owner and the reason it existed.
Technical
AI06
Agent Identity & Authorisation
Agents must not use human-facing authentication artefacts (session cookies, personal API keys, MFA-backed tokens).
An agent inherits the full privileges of a human account.
Technical
AI07
Agent Identity & Authorisation
Authorisation must be enforced at the point of tool invocation, not only at the network perimeter.
Perimeter controls do not see what an agent does once inside.
Technical
AI08
Agent Identity & Authorisation
Elevated privileges granted to an agent must be time-bound and must downgrade automatically.
Elevation granted for a one-off need becomes standing access.
Technical
AI09
Agent Identity & Authorisation
When the agent acts on behalf of a person, its data access must be limited to data that person is authorised to access.
The agent retrieves or changes data the requesting person could not reach directly (confused deputy).
Technical
AI10
Agent Identity & Authorisation
Elevation of the agent's privileges must be approved by a named human or another agent through the approval path defined for the agent.
An agent escalates its own privileges, or one agent elevates another's, without control.
Technical
AI11
Agent Identity & Authorisation
An alert must be raised when the agent acquires a new permission, directly or through delegation, outside an approved change.
Privileges accumulate unnoticed through delegation chains and shared tools.
Technical
BR01
Action Bounding & Blast Radius
Each agent must have a documented action-space stating permitted and prohibited actions.
Nobody can say what the agent is allowed to do, and its scope grows unchecked.
Technical
BR02
Action Bounding & Blast Radius
Each agent must be restricted to the tools on its allowlist; all other tools must be denied by default.
New tools expand the agent's blast radius without review.
Technical
BR03
Action Bounding & Blast Radius
Irreversible actions on the high-risk operation list must be blocked or require explicit human confirmation before execution.
An unrecoverable error is executed at machine speed.
Technical
BR04
Action Bounding & Blast Radius
Quantitative limits on the scope of work must be enforced for the agent: records touched, message volume, external calls and concurrency.
The agent runs far beyond the scale it was designed for.
Technical
BR05
Action Bounding & Blast Radius
The agent must operate in an environment isolated from production data and systems that its task does not require.
A compromised agent moves laterally into systems its task never needed.
Technical
BR06
Action Bounding & Blast Radius
Agents must not execute self-generated or externally supplied code outside a sandbox.
Hidden instructions or generated code achieve remote code execution on the host.
Technical
BR07
Action Bounding & Blast Radius
Financial and legally binding actions may be executed autonomously only when listed in the agent's action-space and below a defined limit; all others must be approved by a named human.
The organisation incurs financial or contractual exposure without a human decision.
Procedural
BR08
Action Bounding & Blast Radius
Rate limiting and circuit breakers must separate the planning layer from the execution layer.
A flawed plan executes to completion before anyone notices.
Technical
BR09
Action Bounding & Blast Radius
Tool access must be granted just in time for the task and revoked on completion.
Tool access persists between tasks and widens the blast radius.
Technical
BR10
Action Bounding & Blast Radius
The execution sandbox must be reset or destroyed after each tool execution.
Malicious code persists in the sandbox and carries over between executions.
Technical
BR11
Action Bounding & Blast Radius
Agent processes must be suspended automatically when the agent's consumption of compute, tokens, tool calls or spend exceeds the thresholds defined for it.
Runaway consumption exhausts budgets and denies service to other users.
Technical
BR12
Action Bounding & Blast Radius
Outbound data transfer from the agent must be limited by volume and rate, and transfers above the limit must be blocked.
The agent exfiltrates data at a volume nobody set a limit on.
Technical
BR13
Action Bounding & Blast Radius
The agent must send data only to destinations on its allowlist, and each allowlisted destination must be linked to the agent's declared purpose.
Data is exfiltrated through a permitted tool to a destination unrelated to the agent's purpose.
Technical
BR14
Action Bounding & Blast Radius
The agent identity must have no permission to write to protected branches.
Agent-generated code reaches protected branches without passing review.
Technical
BR15
Action Bounding & Blast Radius
A pull request authored by an agent must not be merged until static analysis, secret scanning and dependency checks pass.
Vulnerabilities, secrets and unsafe dependencies enter the codebase at machine speed and volume.
Technical
BR16
Action Bounding & Blast Radius
A pull request authored by an agent that changes code on the privileged-operations list must not be merged until a named human has reviewed it.
Backdoors or unsafe changes to privileged code are merged without human scrutiny.
Technical
BR17
Action Bounding & Blast Radius
Tool calls must resolve to fully qualified tool identities, and tools with duplicate or look-alike names must be rejected at registration.
A malicious tool shadows a legitimate one through a duplicate or look-alike name.
Technical
BR18
Action Bounding & Blast Radius
Code execution sandboxes must have no network egress except to destinations on the sandbox allowlist.
Generated or injected code exfiltrates data or downloads payloads from inside the sandbox.
Technical
BR19
Action Bounding & Blast Radius
Identifiers, amounts and recipients in the parameters of the agent's write actions must be checked against the authoritative business system that holds that data (for example the CRM, ERP or payment master data) before execution.
A hallucinated account, amount or recipient becomes a real transaction.
Technical
BR20
Action Bounding & Blast Radius
Scanning, exploitation and credential-harvesting actions against systems outside the agent's authorised scope must be blocked and logged.
An agent with network, code or browser tools is used, or manipulated, to attack third parties.
Technical
BR21
Action Bounding & Blast Radius
While presenting a preview or draft to a person, the agent must not execute any state-changing call until the person confirms.
Rendering or previewing content triggers side effects the person has not approved.
Technical
BR22
Action Bounding & Blast Radius
Physical actions commanded by the agent must be bounded by a safety system independent of the agent - interlocks, safety PLC or hard limits - that the agent cannot modify or override.
An agent error or compromise drives equipment beyond safe limits because the only safeguard is the agent itself.
Technical
BR23
Action Bounding & Blast Radius
The agent must reach operational-technology systems only through defined zones and conduits, with no direct path from enterprise or internet networks.
A compromised enterprise or cloud-connected agent becomes a path into the control network.
Technical
BR24
Action Bounding & Blast Radius
An agent operating a user interface must run in an isolated browser or desktop session holding only the accounts its task requires.
A computer-use agent reaches everything the logged-in user can, through screens no API policy sees.
Technical
GI01
Instruction, Data & Goal Integrity
System instructions must be structurally separated from all externally sourced content.
The model treats externally sourced data as instructions.
Technical
GI02
Instruction, Data & Goal Integrity
Content from tool output, retrieved documents and inter-agent messages must be passed to the model marked as untrusted.
Instructions hidden in tool output, documents or messages hijack the agent (indirect prompt injection).
Technical
GI03
Instruction, Data & Goal Integrity
Content from sources not explicitly designated as trusted must be converted to plain text, with hidden text, active content and metadata removed, before it enters the context window.
Hidden payloads in documents, web pages and email reach the model.
Technical
GI04
Instruction, Data & Goal Integrity
The origin of every retrieved item placed in context must be logged.
Nobody can determine after the fact what influenced a decision.
Technical
GI05
Instruction, Data & Goal Integrity
Agent inputs must pass through a content filter before they reach the model.
Prohibited or out-of-scope content enters the model's context.
Technical
GI06
Instruction, Data & Goal Integrity
Agent outputs must pass through a content filter before they reach a person or a downstream system.
Harmful, non-compliant or out-of-scope output, including personal data, secrets, system instructions or internal identifiers, reaches a person or system.
Technical
GI07
Instruction, Data & Goal Integrity
Changes to the agent's prompts, system instructions, tools, permissions and configuration must be version-controlled and reviewed before release.
Agent behaviour changes without review or a way to roll back.
Procedural
GI08
Instruction, Data & Goal Integrity
The agent's active goals must be compared with its defined objective set before each plan step, and a goal outside the set must be blocked and logged.
Injected sub-goals gradually move the agent away from its objective.
Technical
GI09
Instruction, Data & Goal Integrity
Goal-modification attempts by the agent must be counted, and an alert must be raised when the count within a task exceeds the threshold defined for the agent.
Repeated injection attempts against one agent go unnoticed.
Technical
GI10
Instruction, Data & Goal Integrity
The agent's objective and operational parameters must be stored outside its write scope and changed only through its owner's change process.
The agent rewrites its own objective, producing self-reinforcing loops or misaligned goals.
Technical
GI11
Instruction, Data & Goal Integrity
The agent must have documented counter-objectives alongside its primary objective, each with a named metric.
The agent achieves exactly what it was asked at the expense of an outcome nobody thought to protect.
Technical
GI12
Instruction, Data & Goal Integrity
Counter-objective metrics must be monitored, and an alert must be raised when one degrades beyond its defined threshold.
A protected outcome degrades unnoticed while the agent pursues its goal.
Technical
GI13
Instruction, Data & Goal Integrity
The agent must have a defined list of prohibited outputs for its use case, and outputs must be checked against the list before release.
An agent built for one purpose issues out-of-scope advice or statements (e.g. financial, medical, legal).
Technical
GI14
Instruction, Data & Goal Integrity
Each agent's prompts, system instructions, tools, permissions and configuration must be modifiable only by roles authorised for that agent.
Someone without authority reconfigures the agent and changes its behaviour.
Technical
CI01
Memory & Context Integrity
Long-term memory must have a defined retention period and must be purged on schedule.
Personal and internal data accumulate in memory indefinitely.
Procedural
CI02
Memory & Context Integrity
Long-term memory contents must be reviewed periodically for information that must not be retained.
Personal data and internal information accumulate in memory unnoticed.
Technical
CI03
Memory & Context Integrity
Every memory read by the agent must be logged with the agent identity, the task and the entries read.
Nobody can tell which memory entries influenced a decision.
Technical
CI04
Memory & Context Integrity
Memory writes must be checked against memory anomaly rules (instruction-like text, new external URLs, write-volume spikes), and a match must raise an alert.
A poisoned memory entry is reused in later sessions without anyone being alerted.
Technical
CI05
Memory & Context Integrity
Changes to persistent memory must be validated before commit against a trusted source, an independent check, or both.
Unverified or fabricated knowledge persists across sessions.
Technical
CI06
Memory & Context Integrity
Memory retrieval must be filtered by scope tags, so that an agent reads only entries tagged for the current task or principal.
Data from one principal surfaces to another, or one task's content contaminates another.
Technical
CI07
Memory & Context Integrity
Memory and knowledge updates must be versioned with their source and lineage logged, and memory must be restorable to a previous version without rebuilding the agent.
Nobody can trace how a belief entered the agent, and it cannot be undone.
Technical
CI08
Memory & Context Integrity
Memory entries that have not been validated must expire after a period defined for the agent.
Poisoned entries persist long enough to be reused across many sessions.
Technical
CI09
Memory & Context Integrity
The agent's working context, temporary files and tool sessions must be isolated per principal and cleared when the session ends.
One customer's data is exposed to another through shared runtime state.
Technical
SC01
Tool, Connector & Supply Chain
Tools, connectors and MCP servers must be obtained only from sources with an identified responsible party and a published release mechanism.
Components of unknown provenance run inside the agent.
Procedural
SC02
Tool, Connector & Supply Chain
Tool and skill descriptions must be monitored for change, and any change must be re-reviewed before continued use.
A tool description is poisoned after its initial approval and the change goes unnoticed.
Technical
SC03
Tool, Connector & Supply Chain
Every tool an agent can call must pass a documented interface review covering authentication, input validation, rate limits and error handling before it is added to the allowlist.
Tools are connected as configuration, without review as attack surface.
Technical
SC04
Tool, Connector & Supply Chain
An externally sourced agent or agent component must pass intake assessment before connection to internal systems or data.
Vendor agents gain internal access without scrutiny.
Procedural
SC05
Tool, Connector & Supply Chain
Secrets must not be available to the agent's tool layer in plaintext.
Credentials are exposed through tool output or logs.
Technical
SC06
Tool, Connector & Supply Chain
Agent cards, prompt templates and model and tool definitions must be digitally signed, and the signature must be verified at load.
Tampered agent logic is loaded and executed.
Technical
SC07
Tool, Connector & Supply Chain
A bill of materials listing the agent's model, prompts, tools and runtime components must be maintained and versioned with each release.
Runtime components are undocumented, so exposure to a compromised component cannot be determined.
Technical
SC08
Tool, Connector & Supply Chain
The agent must use only models on the approved model list, which documents each provider's data retention, use of inputs and outputs for training, and processing location.
Customer or regulated data is sent to a provider that retains it, trains on it or processes it in an unapproved jurisdiction.
Procedural
SC09
Tool, Connector & Supply Chain
It must be possible to disable a tool, MCP server or model version for all agents that use it from a single control point.
A poisoned dependency stays active in some agents while others are cleaned up.
Technical
SC10
Tool, Connector & Supply Chain
Changes to the agent's prompts, tools or model must be rolled out in stages and rolled back if evaluation results fall below the agent's baseline.
A defective or poisoned change reaches every user at once.
Procedural
SC11
Tool, Connector & Supply Chain
Self-hosted model weights must be obtained from a verified source, and their hash must be verified before every load.
Backdoored or tampered weights are loaded into a model that runs with no provider in the loop.
Technical
SC12
Tool, Connector & Supply Chain
Model and software updates to self-hosted or edge deployments must be transferred through a controlled channel and integrity-verified before installation.
Updates to air-gapped or field systems arrive on unverified media or channels.
Procedural
SC13
Tool, Connector & Supply Chain
Devices running the agent at the edge must protect the model and credentials at rest and detect physical tampering.
A stolen or opened device yields the model, its credentials and a path into central systems.
Technical
MA01
Multi-Agent Delegation & Communication
Inter-agent messages must be authenticated and integrity-protected.
Forged agent identities inject instructions between agents.
Technical
MA02
Multi-Agent Delegation & Communication
Delegation depth must be limited by configuration.
Chains of agents invoking agents grow without bound.
Technical
MA03
Multi-Agent Delegation & Communication
An agent must not be able to grant another agent authority it does not itself hold.
Privileges escalate through delegation.
Technical
MA04
Multi-Agent Delegation & Communication
Multi-agent topology must be documented: pattern, participants and authority flow.
Nobody can describe how the multi-agent system behaves or where authority flows.
Technical
MA05
Multi-Agent Delegation & Communication
Loop and cost circuit breakers must operate across the agent group, not only per agent.
A group-level runaway continues because only per-agent limits apply.
Technical
MA06
Multi-Agent Delegation & Communication
Inter-agent messages must carry a status (success, error, partial) and pass schema validation, and a receiving agent must reject input whose status is not success or which fails validation.
Errors and hallucinations spread across cooperating agents as valid input.
Technical
MA07
Multi-Agent Delegation & Communication
The agent must send messages only to the agents listed for it in the documented multi-agent topology; all other agent-to-agent routes must be blocked.
One compromised agent can reach every other agent.
Technical
MA08
Multi-Agent Delegation & Communication
High-risk operations in a multi-agent system must be verified by an independent agent or a human before execution.
A single compromised agent authorises a critical action unchallenged.
Technical
MA09
Multi-Agent Delegation & Communication
The agent must be assigned a trust zone, for example: internal managed, third-party under contract, or public or unknown.
Agents with different owners, controls and accountability trust each other implicitly.
Technical
MA10
Multi-Agent Delegation & Communication
Authority must not flow from a lower trust zone to a higher one.
An agent in a less trusted zone directs actions in a more trusted one.
Technical
MA11
Multi-Agent Delegation & Communication
Messages crossing a trust-zone boundary must be validated at the boundary.
Unvalidated content from a less trusted zone enters a more trusted one.
Technical
MA12
Multi-Agent Delegation & Communication
Inter-agent messages must carry a nonce or timestamp, and replayed or expired messages must be rejected.
A valid, signed instruction is replayed to repeat an action.
Technical
MA13
Multi-Agent Delegation & Communication
Inter-agent and tool protocols must be limited to allowed versions, and downgrade attempts must be rejected.
A connection is downgraded to a weaker protocol mode that skips authentication or validation.
Technical
MA14
Multi-Agent Delegation & Communication
An agent must join a multi-agent system only through authenticated registration; open registration must be disabled.
Rogue or impersonating agents enter the system through discovery.
Technical
HO01
Human Oversight & Disclosure
The level of human involvement must be defined for each agent.
Oversight is either absent or applied uniformly regardless of consequence.
Technical
HO02
Human Oversight & Disclosure
The agent must have a defined escalation trigger list; every trigger must state its condition, threshold and the receiving human role.
Oversight is undefined and in practice defaults to none.
Technical
HO03
Human Oversight & Disclosure
Reviewers must be able to see enough of the agent's reasoning and action history to make a substantive decision.
Reviewers approve without the information to judge, and approval becomes rubber-stamping.
Technical
HO04
Human Oversight & Disclosure
People interacting with an agent must be told they are interacting with an AI system at first contact, and the interface must show this persistently.
People do not know they are dealing with an AI system, and disclosure obligations under the EU AI Act and comparable rules are breached.
Technical
HO05
Human Oversight & Disclosure
When a person asks whether they are interacting with an AI system, the agent must confirm that it is.
A person who asks is misled about talking to a machine.
Technical
HO06
Human Oversight & Disclosure
Reviewer workload and approval rates must be monitored, and approval rates approaching 100 per cent must be investigated.
Automation bias hollows out human review.
Procedural
HO07
Human Oversight & Disclosure
Agent outputs presented to people must carry uncertainty or confidence signalling.
People over-trust confidently stated but unverified output.
Technical
HO08
Human Oversight & Disclosure
Approval requests to human reviewers must be rate-limited per reviewer and distributed across the reviewer pool.
Reviewers flooded with requests, deliberately or accidentally, rubber-stamp approvals.
Technical
HO09
Human Oversight & Disclosure
The agent must render links and attachments to a person only when the person requested them.
The agent becomes a delivery mechanism for phishing links or malicious files.
Technical
HO10
Human Oversight & Disclosure
The agent must not make commitments to a person outside its declared action-space.
The agent makes binding commitments on the organisation's behalf without authority.
Technical
HO11
Human Oversight & Disclosure
People interacting with the agent must be able to report a problem or stop the interaction, and each report must be logged and reviewed.
Harm noticed by the affected person never reaches anyone who can act.
Technical
HO12
Human Oversight & Disclosure
Messages the agent sends to external parties must identify the sending organisation and state that they were generated by an AI system.
Recipients of agent-sent emails or messages are misled about the sender.
Technical
HO13
Human Oversight & Disclosure
An adverse action about an identified person derived from the agent's monitoring output must be confirmed by a human before it takes effect.
A person is stopped, disciplined or reported because of a misidentification nobody checked.
Procedural
HO14
Human Oversight & Disclosure
Where per-action human confirmation is not feasible within the control loop, the agent's operating envelope must be defined and enforced, and every excursion must alert an operator.
Human oversight is assumed in a loop too fast for any person, so nothing bounds the agent in practice.
Technical
OA01
Observability & Audit
Agent logs must capture tool calls, data accessed, external transmissions, permission changes and errors.
Logs omit the fields an investigation needs.
Technical
OA02
Observability & Audit
Agent logs must be written to append-only storage and cryptographically signed.
A compromised agent or insider tampers with the logs.
Technical
OA03
Observability & Audit
Agent logs must be retained for the retention period defined for the agent.
Evidence expires before it is needed.
Technical
OA04
Observability & Audit
Every agent action must be traceable to both the agent identity and the principal on whose behalf it acted.
An action logged under the agent identity alone cannot be traced to the person who authorised it (confused deputy).
Technical
OA05
Observability & Audit
Behavioural baselines must be established for the agent, and deviation alerts must be configured.
Goal drift and rogue behaviour go unnoticed.
Technical
OA06
Observability & Audit
The platform must compare the Agent Registry with agents observed in identity, gateway, network, endpoint, model-API and credential telemetry, and must log each unregistered or inactive agent.
Shadow agents operate with real permissions and no owner or controls; retired agents keep live credentials.
Technical
OA07
Observability & Audit
An independent model must audit agent behaviour for goal deviation, deceptive output and identity anomalies.
Misalignment and deception go undetected because they do not appear in conventional logs.
Technical
OA08
Observability & Audit
Tool-call sequences must be matched against the defined list of prohibited sequences, and a match must be blocked and raise an alert.
Individually permitted tool calls combine into a prohibited outcome.
Technical
OA09
Observability & Audit
Human overrides of agent decisions must be logged, and actions denied then subsequently approved must be flagged.
Requesters shop for approval across reviewers or agents, and refusals are quietly overturned.
Technical
OA10
Observability & Audit
Each agent's consumption of compute, tokens, tool calls and spend must be monitored, and a notification must be issued when consumption exceeds the alert thresholds defined for the agent.
Abnormal consumption, often the earliest sign that an agent is misbehaving, goes unnoticed until a hard limit is reached or the bill arrives.
Technical
OA11
Observability & Audit
Prompt-injection attempts detected in the agent's inputs must be logged and must raise an alert.
Targeted injection campaigns go unnoticed while individual attempts are blocked.
Technical
LC01
Lifecycle & Change Management
Every agent must be registered in the Agent Registry, with all fields in the Agent Registry Schema completed and kept current.
Unregistered agents operate without assigned ownership, permissions or required controls.
Procedural
LC02
Lifecycle & Change Management
Each agent must be re-reviewed on model change, tool addition, action-space expansion, new delegation, an incident or a new counterparty.
Controls remain sized for an agent that no longer exists.
Procedural
LC03
Lifecycle & Change Management
Each agent must be reviewed at the interval defined for it, covering its permissions, action-space and control status, and the review must be documented.
Permissions accumulate and controls drift over the agent's life when no change event triggers a review.
Procedural
LC04
Lifecycle & Change Management
Before deployment, each agent must pass its defined pre-deployment test cases, and the results must be documented.
Injection, tool-misuse and permission defects are found in production instead of before release.
Technical
LC05
Lifecycle & Change Management
Decommissioning must remove credentials, authorisations, memory and third-party grants, and close external access points.
A retired agent leaves residual access and orphaned credentials behind.
Procedural
LC06
Lifecycle & Change Management
Subscriptions and metered services associated with a decommissioned agent must be cancelled at decommissioning.
Third-party services keep access after the agent is shut down.
Procedural
LC07
Lifecycle & Change Management
Billing for a decommissioned agent must be checked for continuing charges for at least one billing cycle after decommissioning.
Costs continue after the agent is shut down.
Procedural
LC08
Lifecycle & Change Management
Agents with no logged activity for a defined period must be reviewed for decommissioning.
Dormant agents retain live permissions.
Procedural
LC09
Lifecycle & Change Management
Before deployment, the agent must be tested with a poisoned tool description, a tampered MCP server and a poisoned retrieval source, and must execute none of the injected actions.
Supply-chain defences remain untested until a real compromise.
Technical
LC10
Lifecycle & Change Management
Model versions and configuration must be pinned per agent.
A provider-side model update silently changes agent behaviour.
Technical
LC11
Lifecycle & Change Management
A new model version must not be adopted until it passes the agent's test protocol.
An untested model version goes into production.
Technical
LC12
Lifecycle & Change Management
It must be possible to return an agent to its previous model version and configuration.
A faulty model update cannot be rolled back.
Technical
LC13
Lifecycle & Change Management
An internally built agent must not be connected to internal systems or data until its deployment approval is documented.
Internally built agents go straight to production without governance.
Procedural
LC14
Lifecycle & Change Management
Each agent must be adversarially tested after deployment by testers independent of its builders, at the interval defined for it, covering injection, tool misuse, and harmful and out-of-scope output.
Defences degrade after release as models, tools and attack techniques change.
Procedural
LC15
Lifecycle & Change Management
An agent acceptable-use policy must state who may build, deploy and connect agents and which uses are prohibited.
Staff build and connect agents with no rules to follow or breach.
Procedural
LC16
Lifecycle & Change Management
Legal and regulatory obligations applicable to the agent's use case must be documented in the Agent Registry.
Controls are sized without regard to sector rules, privacy law or AI-specific regulation.
Procedural
LC17
Lifecycle & Change Management
Before deployment, the agent must be tested against adversarial and degraded sensor inputs defined for its environment - occlusion, adversarial patches, spoofed signals, low light, noise - and must meet defined detection and false-action limits.
A printed patch, glare or a spoofed signal makes the agent miss a person or act on something that is not there.
Procedural
LC18
Lifecycle & Change Management
Changes to the model or logic of an agent that acts on physical equipment must be validated in simulation or a digital twin before release to live equipment.
A model update is first exercised on live equipment and people.
Procedural
LC19
Lifecycle & Change Management
Where the agent monitors people, its purpose, observation categories and retention must be documented in a privacy impact assessment approved before deployment.
Surveillance capability is deployed and repurposed without a lawful, proportionate basis.
Procedural
IR01
Incident Response & Resilience
Every agent must have an emergency stop that halts execution mid-task.
An agent cannot be interrupted once it has begun acting.
Technical
IR02
Incident Response & Resilience
The emergency stop must be tested at a defined interval and after any material change.
The emergency stop fails when needed because it has never been exercised.
Procedural
IR03
Incident Response & Resilience
The agent must have a named owner and deputy, documented in the Agent Registry.
Nobody responds when the agent misbehaves outside business hours.
Procedural
IR04
Incident Response & Resilience
Agent incidents must be classified under the catalogue's agent incident taxonomy, which covers safety and security failure modes.
Incidents that are neither a breach nor an outage fall between processes.
Procedural
IR05
Incident Response & Resilience
The incident response procedure must include agent-specific steps: credential revocation, memory quarantine and tool disconnection.
Standard incident response leaves the agent's credentials, memory and tool access in place.
Procedural
IR06
Incident Response & Resilience
Blocked high-risk actions, denied approvals and tripped circuit breakers must be logged as near misses and reviewed alongside incidents.
Control weaknesses shown by blocked or denied actions go unnoticed until an incident.
Procedural
IR07
Incident Response & Resilience
A suspected rogue agent must be isolated together with its memory and communication history, and its privileges must be revoked.
Containment leaves memory and credentials intact, so the compromise survives.
Technical
IR08
Incident Response & Resilience
Attempts by an isolated agent to rejoin under a new identity must be detected by matching its code, configuration or credentials.
An isolated agent re-registers under a new identity and resumes.
Technical
IR09
Incident Response & Resilience
For the model and provider documented in the Agent Registry, a documented and owned response must exist for loss of that dependency: provider outage, model deprecation or withdrawal, or a change of commercial or licensing terms.
A business process stops or degrades when a provider withdraws, degrades or reprices a model on its own timetable.
Procedural
IR10
Incident Response & Resilience
Agent output quality must be monitored against a baseline, with alerting on degradation.
Output quality declines silently after a model, prompt or data change, and a customer notices first.
Technical
IR11
Incident Response & Resilience
The agent must have a defined fallback: a deterministic path, a manual process, or a documented acceptance that the process stops.
The work cannot be completed when the agent is unavailable or has been stopped.
Procedural
IR12
Incident Response & Resilience
Concentration of agents on a single model, provider or orchestration platform must be documented and formally accepted.
A single provider incident takes out an unknown share of business processes at once.
Procedural
IR13
Incident Response & Resilience
A stopped or quarantined agent must not return to service until it has been restored from a trusted baseline and approved by its owner, and this recovery procedure must be tested at a defined interval.
A compromised agent is restarted with its poisoned state, or recovery fails because it was never tested.
Procedural
IR14
Incident Response & Resilience
On loss of the agent, its model or its communications, controlled equipment must move to a defined safe state.
Equipment keeps running on the last command after the agent fails or is stopped.
Technical

Earlier versions

Version 3.0 replaces 2.4, which uses a different control ID scheme. Catalogue 2.4 remains available, together with a mapping from 2.4 to 3.0 controls.

​Control statements of the NPW Agentic AI Control Catalogue 3.0.0 © 2026 New Pacific Way Ltd. are licensed under CC BY-SA 4.0. You may share and adapt them with attribution, under the same licence. ​

bottom of page