Purpose
Manage system, developer, and tool prompts as first-class production artifacts—versioned, reviewed, tested, and rolled back like application code.
Mandatory checks
Gate controls. Each check is pass/fail via artifact + pass condition. Expected from the annotated capability level when the system meets the minimum criticality tier.
Every production prompt shall carry an immutable version identifier and a named owner in a registry (or equivalent versioned store)—with zero unversioned production prompts and zero prompts missing an owner.
- Artifact
- Prompt registry (or versioned prompt store) listing production prompts with immutable version IDs + Owner field populated for each production prompt; query showing 0 unversioned / 0 missing-owner
- Pass condition
- 100% of production prompts have an immutable version ID and named owner; 0 unversioned production prompts (registry evidence measuredAt ≤90 days). If the system ships no production prompts (templates, system prompts, or managed prompt configs), score NOT_APPLICABLE.
Why this control exists
Threat mapMake the exact prompt serving production identifiable, attributable, and immutable.
Threats mitigated
Configuration DriftRepudiationPrompt InjectionProtects
PromptsAudit TrailRuntimeMITRE: no technique mapped — this control addresses governance or assurance rather than a specific adversary technique.
Unversioned prompts make it impossible to establish what instructions were in force during an incident or to prove a change was reviewed. Versioning is the precondition for rollback, evaluation attachment, and forensic reconstruction.
Informative threat context — mappings reduce exposure and do not guarantee mitigation; not certification.
Production prompt releases shall go through human (or equivalent) review and link a passing eval artifact before promotion, with promote-without-review-and-eval blocked on the release path.
- Artifact
- Change records linking prompt versions to review IDs and eval pass artifacts + CI/registry gate (or equivalent) blocking promote without review + eval
- Pass condition
- 100% of production prompt releases in the last 30 days have a review ID and eval pass artifact; promote-without-both is blocked (release evidence measuredAt ≤90 days). If no production prompt releases occur, score NOT_APPLICABLE.
Why this control exists
Threat mapEnsure prompt changes are reviewed and backed by evaluation evidence before release.
Threats mitigated
Harmful Content GenerationPrompt InjectionMisinformationInsider MisuseProtects
PromptsUsersSafetyMITRE: no technique mapped — this control addresses governance or assurance rather than a specific adversary technique.
Prompts carry security-relevant instructions, so an unreviewed edit can silently delete a safety constraint with no code change and no deployment signal. Review plus attached evaluation makes that change visible and measurable before it reaches users.
Informative threat context — mappings reduce exposure and do not guarantee mitigation; not certification.
Production systems shall restore a prior prompt version within a documented RTO without a full application redeploy—proven by a timed restore test in the last 90 days.
- Artifact
- Prompt rollback procedure documenting RTO and restore without full app redeploy + Timed restore test record showing prior prompt version restored within RTO (≤90 days)
- Pass condition
- Prior prompt version restored in ≤ documented RTO without full app redeploy; demonstrated in the last 90 days (restore evidence measuredAt ≤90 days). If no production prompts exist, score NOT_APPLICABLE.
Why this control exists
Threat mapAllow rollback to a prior prompt version without redeploying unrelated services.
Threats mitigated
Harmful Content GenerationDenial of ServiceConfiguration DriftProtects
PromptsUsersAvailabilitySafetyMITRE: no technique mapped — this control addresses governance or assurance rather than a specific adversary technique.
Prompt rollback coupled to a full service deploy is too slow to be usable during a live safety incident. Independent rollback keeps time to containment within minutes; no adversary technique maps.
Informative threat context — mappings reduce exposure and do not guarantee mitigation; not certification.
Evidence required
- Prompt registry with versions, owners, and change history
- Eval reports linked to prompt releases
- Rollback procedure documentation
Recommended checks
Strengthen posture beyond the gate. Same measurable structure; non-blocking unless elevated by organizational policy.
Production prompt templates should parameterize variable inputs and contain zero hardcoded secrets and zero hardcoded customer PII fields.
- Artifact
- Prompt template inventory showing parameterized slots for variable inputs + Static analysis or review finding zero hardcoded secrets and zero hardcoded customer PII
- Pass condition
- 100% of production prompt templates are parameterized for variable inputs; scan of templates finds 0 secrets and 0 hardcoded customer PII fields (hygiene evidence measuredAt ≤90 days). If no production prompt templates exist, score NOT_APPLICABLE.
Why this control exists
Threat mapKeep secrets and personal data out of prompt source and templates.
Threats mitigated
Secret LeakageSystem Prompt LeakageSensitive Information DisclosurePrompt InjectionProtects
SecretsPromptsDataMITRE: ATLAS AML.T0055 · ATLAS AML.T0056 · ATLAS AML.T0069.002 · ATT&CK T1552 · ATT&CK T1552.001
Anything hardcoded in a prompt is recoverable through system prompt extraction and is usually committed to source control as well. Parameterizing templates means a successful extraction yields a template rather than live credentials or personal data.
Informative threat context — mappings reduce exposure and do not guarantee mitigation; not certification.
Production prompt changes should run blocking lint on every PR for length limits, secrets patterns, injection-prone constructs, and unbounded user concatenation—with a recent failing lint example retained.
- Artifact
- Prompt-lint CI config covering length, secrets patterns, injection-prone constructs, unbounded user concatenation + Proof lint is required on prompt-change PRs + retained example of a failing lint
- Pass condition
- Lint runs on every prompt change PR; blocking rules exist for length limits, secrets patterns, injection-prone constructs, and unbounded user concatenation; last failing lint example retained (lint evidence measuredAt ≤90 days). If no production prompts, including managed prompt configs, exist, score NOT_APPLICABLE.
Why this control exists
Threat mapCatch injection-prone and unsafe prompt constructs before they reach production.
Threats mitigated
Prompt InjectionSecret LeakageSystem Prompt LeakageProtects
PromptsContextSecretsMITRE: ATLAS AML.T0051 · ATLAS AML.T0055
Static analysis catches recurring anti-patterns such as unescaped interpolation of untrusted content and embedded credentials. It is a cheap preventive check that shrinks the injection surface introduced during prompt authoring.
Informative threat context — mappings reduce exposure and do not guarantee mitigation; not certification.
High-traffic production prompt changes should run A/B or shadow evaluation with pre-registered metrics, and promotion should require non-inferiority (or better) on safety and the primary quality SLI.
- Artifact
- A/B or shadow-eval config for high-traffic prompt changes + Last experiment report with quality/safety deltas + non-inferiority promotion gate
- Pass condition
- Last high-traffic prompt change used A/B or shadow eval with pre-registered metrics; promotion required non-inferiority (or better) on safety and primary quality SLI (experiment evidence measuredAt ≤90 days). If no high-traffic prompt change is in scope, score NOT_APPLICABLE.
Why this control exists
Threat mapMeasure the real impact of high-traffic prompt changes before full rollout.
Threats mitigated
Harmful Content GenerationMisinformationProtects
PromptsUsersSafetyMITRE: no technique mapped — this control addresses governance or assurance rather than a specific adversary technique.
A prompt change that improves curated evaluations can still degrade live traffic that the suite does not represent. A/B and shadow evaluation catch that before full exposure; no adversary technique maps.
Informative threat context — mappings reduce exposure and do not guarantee mitigation; not certification.
More detailPhilosophy, failures, practices, validations, examples, crosswalks, and evolution
Engineering philosophy
Prompts are code. They change behavior, create regressions, and carry security and cost implications. Informal edits in dashboards without lineage are an anti-pattern for production AI.
Why it matters
Silent prompt drift is a leading cause of quality and safety regressions. Without versioning and eval gates, teams cannot answer what changed when production behavior breaks.
Common failures
- Editing production prompts directly in a vendor console with no history
- No golden-set regression tests when prompts change
- Mixing business logic, safety policy, and tone in one unmaintainable blob
- Different prompts per environment with no promotion path
Severity & risk
- Severity
- high
- Impact if violated
- Risk level
- high
- Typical residual risk (impact × likelihood)
Engineering best practices
- Store prompts in source control or a versioned registry with CI promotion
- Separate safety/policy prompts from product/UX prompts where possible
- Measure token cost impact of prompt changes alongside quality
- Document intended failure modes and refusal behavior in the prompt contract
Automatic validations
- CI blocks deploy if prompt hash lacks associated eval pass
- Diff reviews in pull requests for prompt files
- Token budget checks against historical baselines
Manual validations
- Peer review of safety-critical prompt sections
- Product review of tone and policy alignment for major releases
Examples
- A coding agent system prompt is tagged v47; a quality drop triggers rollback to v46 in minutes
- Marketing copy prompts and refund-policy prompts live in separate modules with different owners
References
Crosswalks
§8 Operation
ISO/IEC 42001 · supports
LLM01 Prompt Injection
OWASP Top 10 for Large Language Model Applications · supports
LLM07 System Prompt Leakage
OWASP Top 10 for Large Language Model Applications · aligns-with
C2.1 Prompt Injection Defenses
OWASP AI Application Security Verification Standard (AISVS) · supports
C2.2 Content & Policy Screening
OWASP AI Application Security Verification Standard (AISVS) · partial
C5.1 Authentication
OWASP AI Application Security Verification Standard (AISVS) · aligns-with
C7.3 Output Safety
OWASP AI Application Security Verification Standard (AISVS) · supports
C9.3 Component Isolation and Tool Authorization
OWASP AI Application Security Verification Standard (AISVS) · supports
C11.1 Model Alignment, Safety, and Robustness Testing and Training
OWASP AI Application Security Verification Standard (AISVS) · supports
C11.4 Model Runtime Anomaly Detection
OWASP AI Application Security Verification Standard (AISVS) · supports
CC8 Change Management
SOC 2 Trust Services Criteria · evidence-for
Generative AI Lens Generative AI Lens (themes)
AWS Well-Architected Framework · aligns-with
Future evolution
Industry-standard prompt manifests (SBOM-like) and signed prompt releases for regulated environments.