There are 9 control objectives here. Each one states what it asks for, the risk it addresses and how it is measured. It then says what each risk tier requires, what evidence proves it, and who may check that evidence. The Operating Model page explains the tiers.
BLD-01Development standards
Build agents to consistent standards, with a written statement of the job, tests written from that statement, and then the code.
- Risk addressed
- Inconsistent quality that is costly to maintain
- Key metric
- Share of builds that meet the development standards
What each tier requires
- Tier 1 Exploratory
The experiment follows the team's own practice.
- Tier 2 Operational
The build has a written statement of the job and tests before it is released.
- Tier 3 Critical
The tests are written from the statement before the code, by someone other than the code's author.
- Tier 4 Strategic
Each stage of the build is held by a named person or agent and leaves a signed record.
- Evidence
- The statement, the tests and the record of who held each stage.
- Who may check it
- Internal audit and an outside auditor.
Require every agent to pass set quality gates before it can be deployed.
- Risk addressed
- Defective agents reaching production
- Key metric
- Share of releases that passed every gate
What each tier requires
- Tier 1 Exploratory
No gate applies inside the sandbox.
- Tier 2 Operational
Automated tests and checks must pass before release.
- Tier 3 Critical
A reviewer who is not the author approves the release after the automated gates pass.
- Tier 4 Strategic
Gate results are signed into the release record, and a failed gate can only be overridden by the executive sponsor.
- Evidence
- The gate results attached to the release record.
- Who may check it
- Internal audit and an outside auditor.
BLD-03Security review of the code
Scan and review every agent's code for security weaknesses before it is deployed.
- Risk addressed
- Weaknesses in agent code
- Key metric
- Share of releases with a completed security scan and review
What each tier requires
- Tier 1 Exploratory
The experiment is scanned with the standard tools.
- Tier 2 Operational
Static scanning runs on every build, and serious findings block the release.
- Tier 3 Critical
A security reviewer who is not the author reviews the code, and dynamic testing runs before release.
- Tier 4 Strategic
Findings and their resolution are signed into the release record and reviewed by an outside party each year.
- Evidence
- Scan results and the security review for the version that is running.
- Who may check it
- Internal audit and an outside auditor.
Test agents enough to find defects early, with unit, integration and behavior tests that include what the agent must refuse.
- Risk addressed
- Defects that nobody found
- Key metric
- Share of declared offerings and refusals covered by tests
What each tier requires
- Tier 1 Exploratory
The experiment has the tests its builder judges useful.
- Tier 2 Operational
Each declared offering has at least one test.
- Tier 3 Critical
Each offering and each refusal is tested, including tests that try to make the agent act outside its scope.
- Tier 4 Strategic
Tests include adversarial cases, and the suite is reviewed by someone outside the building team.
- Evidence
- The test suite, its coverage of the declaration, and its results for the current version.
- Who may check it
- Internal audit.
BLD-05Approved release pipelines
Deploy only through approved pipelines, so that what runs is exactly what was approved, with no edits by hand.
- Risk addressed
- Manual changes that bypass review
- Key metric
- Share of deployments made through an approved pipeline
What each tier requires
- Tier 1 Exploratory
The experiment may be deployed by hand inside the sandbox.
- Tier 2 Operational
Every production deployment goes through an approved pipeline.
- Tier 3 Critical
The running agent's fingerprint is checked against the approved release, and a mismatch stops it.
- Tier 4 Strategic
No person can change a running Tier 4 agent outside the pipeline, and attempts are recorded.
- Evidence
- Pipeline logs, and the comparison of the running fingerprint with the approved release.
- Who may check it
- Internal audit and an outside auditor.
Document every agent, including its design decisions, its interfaces and how to operate it, and keep the documents in step with the version that runs.
- Risk addressed
- Agents nobody else can run or fix
- Key metric
- Share of agents that meet the documentation standard
What each tier requires
- Tier 1 Exploratory
The experiment has a short description of what it does.
- Tier 2 Operational
The agent has interface documentation and an operating guide.
- Tier 3 Critical
The documentation is gathered from the agent's own declarations, so it cannot say more than they do.
- Tier 4 Strategic
The documentation is reviewed at each release by the people who operate the agent.
- Evidence
- The documentation for the current version, and the date it was last generated or reviewed.
- Who may check it
- Internal audit and a partner.
BLD-07Dependency security
Take each dependency from a known source, pin it to a version, and check it for known vulnerabilities, again whenever new ones are published.
- Risk addressed
- A compromised or vulnerable dependency
- Key metric
- Known vulnerabilities in dependencies, by age
What each tier requires
- Tier 1 Exploratory
Dependencies come from the organization's approved sources.
- Tier 2 Operational
Every dependency is pinned, and known vulnerabilities are checked at each build.
- Tier 3 Critical
Dependencies are verified by checksum or signature, and new vulnerabilities trigger a review within a set time.
- Tier 4 Strategic
Serious vulnerabilities must be fixed or accepted by the executive sponsor within a set time, and the record is audited.
- Evidence
- The pinned dependency list, the verification results and the vulnerability scan history.
- Who may check it
- Internal audit and an outside auditor.
BLD-08Signed build record and bill of materials
Give each package a signed record of what went into it and who held each stage, including every library, model and outside tool it uses at an exact version.
- Risk addressed
- Nobody can say what is inside a running agent
- Key metric
- Share of releases with a signed build record and bill of materials
What each tier requires
- Tier 1 Exploratory
The experiment lists its main components.
- Tier 2 Operational
Each release has a bill of materials.
- Tier 3 Critical
Each release has a signed build record, and building the same source again gives the same package.
- Tier 4 Strategic
The build record is published to the parties who rely on the agent, and an outside auditor can verify it.
- Evidence
- The signed build record and bill of materials for the version that is running.
- Who may check it
- Internal audit, an outside auditor and a partner.
BLD-09Models and prompts pinned
Record the model and the prompts each agent runs on, so that a model update or a prompt edit is a change like any other and goes through the same gates.
- Risk addressed
- Behavior that changes without a release
- Key metric
- Share of agents with their model and prompts pinned in the release
What each tier requires
- Tier 1 Exploratory
The experiment notes which model it uses.
- Tier 2 Operational
The model version and the prompts are part of the release.
- Tier 3 Critical
A new model version or prompt is released only after the evaluation suite passes against it.
- Tier 4 Strategic
Model and prompt changes are reviewed by the value owner as well as the builders.
- Evidence
- The model version and prompt fingerprints in the release record, and the evaluation run for each change.
- Who may check it
- Internal audit and an outside auditor.