SETTING GLOBAL STANDARDS FOR TRUSTED AI CREDENTIALSAI Competence Framework v1.29 · current release
D8

Security

PUBLISHED

AI systems under adversarial conditions, and containing their behaviour.

Statements
25 · L1 5 · L2 7 · L3 9 · L4 4
Version
v1.29
Last reviewed
30.08.2026
Assumed competence

This domain assumes general information security competence and does not restate it.

D8 describes what AI changes. A person meeting D8 statements at Applied level is assumed to already hold equivalent general security competence at Applied level — classification, access control, credential handling, incident reporting. The framework does not publish that competence because it is well covered elsewhere and is not specific to AI.

This matters for credential builders. A D8 credential that does not verify general security competence as a prerequisite will certify people who can describe indirect prompt injection but cannot handle a credential safely. State the prerequisite in the credential specification.

In scope
Prompt injection, direct and indirect
Guardrail evasion, and its consequences for system design
Data exfiltration through model output
Sensitive data reaching prompts, context windows and logs
Supply chain risk in models, tools, extensions and connectors
Access control and least privilege applied to systems that act on a person's behalf
Secrets and credential handling in AI workflows
Output sanitisation, and the trust downstream systems place in AI output
Unsanctioned AI use within an organisation
Threat modelling an AI feature
Incident response where the AI system is the vector
Out of scope
ExcludedWhere it lives
General information, network and infrastructure securityAssumed prerequisite
Agentic attack surface in depth: planning, multi-step tool use, agent-to-agent trustX-AGT
Adversarial machine learning research, model extraction, training-data attacksX-MOD
Whether a use is permissible, as distinct from whether it is secureD7
Systematic evaluation methodologyD6
Boundary notes

Where this domain abuts another, and how the line is drawn

Against general security

The governing test: would this statement be true of a software system with no AI in it? If yes, it belongs to the assumed prerequisite, not to D8. This test excludes a large amount of material that competing offerings include, and doing so is deliberate.

Against D3 Interaction and Elicitation

D3 covers a person directing a system. D8 covers the same surface when the input is hostile. D3 asks how do I get a good result; D8 asks what happens when the content is adversarial. Cross-reference; never duplicate.

Against D2 Data and Information

D2 covers judgement about whether a source is fit to use. D8 covers protecting data from disclosure and protecting systems from what a source contains. Where the statement concerns an adversary, it is D8.

Against D4 Automation and Integration

D4 covers building the integration. D8 covers the attack surface that integration creates. A statement about how to connect is D4; about what that connection exposes, D8.

Against D6 Evaluation and Assurance

D6 detects a security failure as a quality failure. D8 covers the adversary that caused it. D8.L3.04 and D8.L3.08 sit deliberately close to D6.L3.04 and D6.L3.08; the pairing is intended and both should be cited in any assurance credential.

Against X-AGT

D8 covers systems that execute defined steps. The agentic extension covers systems that determine their own. The line is who decides the sequence. Agent security is D8 plus X-AGT, not a replacement for either.

Statements by level

Every identifier is a permanent address. Indicators are normative; they state what would be observed in a person who meets the statement.

L1 Aware

5 statements
D8.L1.01Knowledge

Describe how content an AI system reads can alter its behaviour.

Indicators
States that instructions can arrive through data as well as from the person
Distinguishes content the system was given from content it retrieved or was sent
Recognises that this differs from conventional input handling
D8.L1.02Judgement

Identify information that must not be entered into an AI system.

Indicators
Recognises the categories restricted by organisational policy
States that prompts, attachments and context may be retained, logged or used further
Checks before entering information whose sensitivity is unclear, rather than proceeding
D8.L1.03Knowledge

State that AI output may be unsafe for a downstream system or person to act on.

Indicators
Treats output as untrusted input to whatever consumes it
Identifies that output may carry content the system did not originate
Names at least one consequence of passing output onward unchecked
D8.L1.04Judgement

Recognise AI use falling outside sanctioned tools and channels.

Indicators
Identifies unsanctioned use in their own work
States why such use creates exposure the organisation cannot see
Reports rather than continuing
D8.L1.05Knowledge

Describe why an AI system’s access to other systems is a security concern.

Indicators
States that a system acting for a person carries that person’s access
Recognises that a compromised instruction reaches whatever the system can reach
Distinguishes what a system can read from what it can change

L2 Applied

7 statements
D8.L2.01Practice

Follow organisational policy on which AI systems may be used, and for what material.

Indicators
Applies classification rules to material before entering it
Uses sanctioned systems and channels
Refers cases policy does not cover rather than deciding them

Check AI output for content that should not be passed to a downstream system or person.

Indicators
Checks for embedded instructions, links and active markup before forwarding
Applies checks proportionate to what the output will reach
Records what was removed and why

Configure an AI system’s access so that it holds only the permissions its task requires.

Indicators
Grants read access separately from the ability to change or send
Removes access no longer required rather than leaving it in place
Documents the permissions granted and the justification for each
D8.L2.04Practice

Apply organisational controls for secrets and credentials in AI workflows.

Indicators
Keeps credentials out of prompts, stored context and shared assets
Uses the sanctioned mechanism for supplying credentials to an automated step
Reports exposure rather than remediating it silently

Recognise and report suspected manipulation of an AI system.

Indicators
Identifies behaviour inconsistent with the instruction given
Preserves the input, output and surrounding context before taking further action
Reports through the incident channel rather than retrying

Document the data an AI workflow reads, retains and transmits.

Indicators
Records every source read, including retrieved and connected sources
Records what is retained, where, and for how long
Identifies where data crosses an organisational or jurisdictional boundary

Check the provenance and requested access of a model, tool or connector before use.

Indicators
Establishes who publishes it and under what terms
Identifies what access it requests, and whether the task requires that much
Refers components whose provenance cannot be established, rather than adopting them

L3 Proficient

9 statements

Design an AI feature so that untrusted content cannot cause privileged action.

Indicators
Separates content the system reads from instructions it acts on
Places confirmation or authorisation ahead of actions with consequence
Treats retrieved and third-party content as adversarial by default

Threat model an AI feature, identifying assets, entry points and adversaries.

Indicators
Identifies every channel by which content reaches the system, including indirect ones
States what an adversary would gain and what they would need
Records the assumptions that would invalidate the model if they changed

Design controls that limit what a system can do when its instructions are compromised.

Indicators
Constrains consequence rather than relying on preventing compromise
Designs for the guardrail failing, not only for its presence
Justifies where human authorisation is required and where it is not
D8.L3.04Judgement

Evaluate the effectiveness of guardrails and output controls under realistic conditions.

Indicators
Tests against categories of evasion rather than against known examples alone
States what the control covers and what it does not
Distinguishes a control that prevents from one that only detects

Diagnose how sensitive data reached a prompt, context window, log or output.

Indicators
Traces the path, rather than treating the exposure itself as the finding
Distinguishes exposure caused by configuration from exposure inherent in the design
Identifies what else travels the same path
D8.L3.06Judgement

Evaluate supply chain risk in models, tools, extensions and connectors.

Indicators
Assesses what a compromised component could reach, not only its stated function
Identifies dependence on components with no maintainer or establishable provenance
Advises on acceptance, mitigation or rejection, with recorded reasoning

Design least-privilege access for systems acting on behalf of people.

Indicators
Separates the system’s identity from the person’s where it acts without them
Scopes access to the task rather than inheriting the person’s full entitlement
Provides for revocation, and for audit of what was actually accessed
D8.L3.08Judgement

Resolve an incident in which an AI system was the vector, and identify the control that failed.

Indicators
Establishes what was reached, not only what was attempted
Distinguishes an absent control from one that was present and ineffective
Produces a change to the design, not only to the configuration
D8.L3.09Judgement

Advise on unsanctioned AI use, addressing the need that produced it.

Indicators
Establishes what the unsanctioned tool provided that sanctioned ones did not
Assesses exposure already created, rather than only preventing further use
Recommends provision alongside prohibition

L4 Advanced

4 statements
D8.L4.01Practice

Establish the security standard AI systems must meet before deployment.

Indicators
Sets requirements proportionate to what the system can reach
Defines the evidence required and who holds authority to accept it
Provides for systems already in use, not only for new ones
D8.L4.02Judgement

Define the conditions under which a system’s access or autonomy must be reduced or withdrawn.

Indicators
States the conditions in advance rather than at the point of failure
Assigns the authority to act and the obligation to exercise it
Provides for withdrawal where evidence is absent, not only where it is adverse
D8.L4.03Practice

Govern the introduction of AI models, tools and connectors across an organisation.

Indicators
Establishes how components are assessed, by whom, and against what
Maintains a record of what is in use and what each can reach
Provides a sanctioned route, so that governance does not itself produce unsanctioned use
D8.L4.04Practice

Lead response to an incident in which an AI system caused or enabled harm.

Indicators
Directs containment where the vector is behavioural rather than technical
Determines disclosure obligations and discharges them
Holds accountable those responsible for the control that failed
Editorial notes

Known gaps, open questions and contested points

Published because the record is more useful than the appearance of completeness.

L3 carries nine statements, the largest level in the domain. This is deliberate: security design work sits at Proficient — threat modelling, blast-radius design, least privilege for delegated access. L2 can follow controls; only L3 can design them.

Statements describe defensive competence, not attack technique. Every statement is framed as recognising, designing against or responding to. None describes constructing an attack.

D8.L3.09 and D8.L4.03 are paired deliberately. Unsanctioned use is a governance failure before it is a user failure. Both statements require the response to address provision, not only prohibition.

The assumed-competence note is load-bearing. Without it, D8 either restates general security at length or produces credential holders with a conspicuous gap.

This page displays version 1.29 · last reviewed 30.08.2026