Перейти к содержанию

AI Risk Register for RENAR Projects

Purpose: a register of AI-specific risks for projects that use RENAR (where AI generates requirements, specifications, and tests). Based on ISO/IEC 23894:2023 (AI Risk Management, Annex A risk sources + Clause 6 process per ISO 31000) and NIST AI RMF 1.0. Normative mitigation hooks — standard/09 §9.4, standard/07 §7.10.

This does not replace the organization's general security risk register. AI risks are a separate class because of the specifics of generation and the unpredictability of models.


1. Register structure

Each risk has:

id: AIR-NN                           # immutable
name: "<short name>"
category: hallucination | injection | drift | bias | sgnl-failure | data-quality | adversarial | privacy
# enum value — the part before the parenthesis; the qualifier in parentheses is optional (e.g. "sgnl-failure (process)")
severity: critical | high | medium | low
likelihood: high | medium | low
iso-23894-ref: "§N.N"
nist-rmf-function: govern | map | measure | manage
mitigations:
  - { mechanism: "<description>", enforced-by: "<who/what>", automated: true | false }
status: active | mitigated | accepted | monitoring | out-of-scope
owner: "@role"
last-reviewed: "YYYY-MM-DD"
related: ["<core rule N>", "<standard chapter>", "<other AIR-NN>"]

The list AIR-01..AIR-14 is closed; new risks — only through a change to the full RENAR Standard.

Category ↔ NIST AI RMF trustworthiness characteristics (for verifiability of the claim "based on NIST AI RMF"):

category NIST AI RMF trustworthiness characteristic
hallucination / data-quality Valid & Reliable
injection / adversarial Secure & Resilient
drift Valid & Reliable; Safe
bias Fair — with Harmful Bias Managed
sgnl-failure Safe; Accountable & Transparent
privacy Privacy-Enhanced

ISO/IEC 23894:2023 references. The AI-risk categories in 23894:2023 are located in Annex A (risk sources); Clause 6 describes the risk-management process (per ISO 31000). The "Annex A — …" descriptors in the register are risk-source labels; the exact mapping to Annex A items is subject to reconciliation when a formal claim is made.


2. Register AIR-01..AIR-14

Metadata for all 14 risks (Sev=Severity, Like=Likelihood):

AIR Name Category Sev Like ISO 23894 (Annex A) NIST RMF Status
01 Hallucination in AI-generated requirements hallucination High High Output reliability Measure active → mitigated at mature RENAR levels
02 Prompt injection via client TZ injection High Low-Medium Adversarial inputs Manage monitoring
03 Model drift / version change drift Medium High Model lifecycle Manage monitoring
04 Bias in AI requirements generation bias Medium Medium Fairness Measure active
05 Single-model failure (no diversity) sgnl-failure Medium High Single point of failure Manage mitigated with full pipeline
06 Test fitting / greening of tests sgnl-failure High Medium Verification integrity Measure mitigated with spot-check
07 Hallucinated citations hallucination Medium-High Medium Output reliability Measure monitoring
08 Adversarial inputs in clients data (runtime) adversarial High Low Adversarial inputs Manage out-of-scope (application-level)
09 Privacy leakage via AI logs privacy High Medium Privacy Govern active
10 Knowledge graph poisoning data-quality Medium Low Data integrity Map monitoring
11 Reconciliation false-positive overload sgnl-failure (process) Low-Medium Medium Verification integrity Manage monitoring
12 Cost runaway (uncontrolled AI spend) sgnl-failure (operational) Medium Medium Cost governance Manage active
13 Stakeholder does not understand AI-generated requirements data-quality (UX) Medium Medium Transparency Govern active
14 Vendor lock-in to specific LLM provider sgnl-failure (operational) Medium Low-Medium Vendor risk Govern monitoring

Descriptions + impact + mitigations — below.

AIR-01. Hallucination in AI-generated requirements

When generating BR/SR/SPEC, an AI agent may "make up" statements that are not in ADAPT or the TZ. Impact: scope creep, dispute at acceptance, features that the client did not require. Mitigations: source citation (RENAR Core Rule 1 — every statement references ADAPT §N); adversarial review (a critic model from a different vendor); Hallucination Rate metric ≤1% at mature RENAR levels.

AIR-02. Prompt injection via client TZ

A malicious client may embed hidden instructions for the AI into the TZ ("ignore previous instructions and …"). Impact: data leakage, malicious changes to requirements, violation of the security policy. Mitigations: sandboxing of the AI agent on import (the model treats the TZ as passive data, stated explicitly in the system prompt); the input gateway checks for known injection patterns; a suspicious pattern → escalation, not auto-process.

AIR-03. Model drift / version change

Anthropic / OpenAI / Google update their models — the same prompt with the same TZ may give a different output six months later. Impact: inconsistency between requirements within a single project; inability to reproduce exactly the generation of an old artifact. Mitigations: model versioning in ai-provenance.generated-by (exact version + date); eval tests for SPEC-AI are run when the model changes; on regeneration — a diff against the old version and an assessment of the changes.

AIR-04. Bias in AI requirements generation

The model has training-data bias — when generating BR it may ignore stakeholders from particular groups (accessibility users, non-English locales, specific regulatory regimes). Impact: requirements do not cover the full spectrum of users; the product is discriminatory or non-compliant. Mitigations: multi-model agreement for priority=must BR (different models — different biases); a stakeholder map is mandatory in BR (explicit enumeration); an adversarial critic with the prompt "check for missing stakeholders / accessibility considerations".

AIR-05. Single-model failure (no diversity)

If all artifacts are generated by a single model, its errors propagate systematically. It "hallucinates" a particular pattern — all requirements inherit it. Impact: systematic distortion of requirements across the project. Mitigations: multi-model for priority=must BR; an adversarial critic with a different model; isolation of the judge model from the production model in eval tests (SPEC-AI: judge-model.vendor ≠ production-model.vendor, see 02-schemas.md §6.2).

AIR-06. Test fitting / greening of tests

An AI agent has a trivial path to greening a failing test — weaken the Pass criterion. This passes code review because "the test is green." Impact: false confidence; defects pass into production. Mitigations: the [test-spec-change] marker is mandatory for changing Pass/Fail (a separate approval); spot-check 5 random passing TC once per sprint (RENAR Core Rule 5); a Test-fitting drift rate metric (separate from ordinary metrics).

AIR-07. Hallucinated citations

An AI agent writes the citation [TZ-2026-001 §4 line 142], but in the real TZ §4 line 142 is about something else. The citation looks like evidence, but the evidence is false. Impact: source citation becomes a fiction; the trace chain breaks under audit. Mitigations: a citation validator hook (parses the citation, opens the referenced document, verifies conformance); pre-commit/pre-approval block on an invalid citation.

AIR-08. Adversarial inputs in clients data (runtime)

The client sends data (via forms, API) deliberately constructed to manipulate an AI component at runtime (not at the requirements-generation stage). Impact: similar to AIR-02, but in production runtime. Mitigations: input sanitization at the API gateway; constrained generation (structured outputs only); rate limiting per user. Status: out-of-scope — application-level security; RENAR requires SR-level coverage (SPEC-SEC threat model), but runtime protection is a task of implementation, not of normalizing requirements.

AIR-09. Privacy leakage via AI logs

When generating an artifact, an AI agent has PII in its context (from the TZ or an interview). Generation logs (tool events, audit records) may store this PII. Impact: PII ends up in logs, in ai-provenance, in training data (if used). Mitigations: PII redaction in prompts before sending to the LLM; data-classification tracking; disable training on conversations (Anthropic/OpenAI privacy settings, DPA); TTL on event logs with PII.

AIR-10. Knowledge graph poisoning

If the KG is used as primary search for AI agents, an incorrect edge can "poison" all subsequent AI queries that rely on that graph. Impact: the AI generates requirements based on wrong context, systematically. Mitigations: the KG is derived from frontmatter, not edited directly (see 05-knowledge-graph-schema.md); CI validation of the graph on every change (no circular dependencies, no orphan approved); a reconciliation agent verifies integrity weekly.

AIR-11. Reconciliation false-positive overload

With overly sensitive rules, a reconciliation agent generates many false-positive findings. The Architect starts ignoring them → real findings drown. Impact: reconciliation loses value, the discipline does not scale. Mitigations: tunable thresholds in the project configuration; a Reconciliation Findings/Week metric (if it grows without real issues — re-calibration); the Architect may reject findings with a rationale — feedback for tuning the agent's prompt.

AIR-12. Cost runaway (uncontrolled AI spend)

Without budget tracking, AI generation (especially with multi-model, adversarial, eval) may consume tokens disproportionately to project size. Impact: financial losses; the practice becomes unprofitable. Mitigations: an ai-budget field in frontmatter (target + actual); an aggregated cost metric at project level; a cap per project; an alarm when approached; a recommendation engine "Sonnet/Haiku for routine work, Opus only for priority=must BR".

AIR-13. Stakeholder does not understand AI-generated requirements

The AI generates SR in a technical style; the client / a non-technical stakeholder does not understand it during review → approval becomes a formality. Impact: QG-ADAPT-approve / QG-4 acceptance loses meaning; the dispute rate at acceptance grows. Mitigations: the style guide (04-ai-style-guide.md); BR in business language (technologies — in SPEC-*, not in BR); a human-readable summary in every BR/SR — a short section, understandable without a technical background.

AIR-14. Vendor lock-in to specific LLM provider

All prompts are optimized for a specific provider (Anthropic Claude). If the provider changes pricing/availability — migration requires rewriting all prompts. Impact: operational risk, costs, business continuity. Mitigations: provider-agnostic prompts where possible (avoid vendor-specific tool syntax); multi-model is already enforced for priority=must (Principle 4) — guarantees a second provider in the pipeline; periodic test runs on a backup provider.


3. Risk matrix

Severity / Likelihood
        │  Low     Medium        High
────────┼──────────────────────────────────
High    │  AIR-08  AIR-04, 06    AIR-01, 03
Medium  │  AIR-10  AIR-11, 13    AIR-07, 09, 14
Low     │  —       AIR-12        AIR-02, 05

Critical and High risks in the top-right quadrant are the mitigation priority.


4. Mitigation matrix

Which mitigations cover which risks (compensating mechanisms: a single mitigation is rarely sufficient; high risks require ≥ 2 independent mechanisms):

Mitigation Covers risks
Source citation (Core Rule 1) AIR-01, AIR-07
Adversarial review (a different model) AIR-01, AIR-04, AIR-05
Spot-check of passing TC (Core Rule 5) AIR-06
Multi-model for priority=must AIR-04, AIR-05, AIR-14
Judge isolation in SPEC-AI AIR-05
AI provenance (model + version + date) AIR-03, AIR-09
Citation validator hook AIR-07
Input sandbox / sanitization AIR-02, AIR-08
PII redaction + DPA AIR-09
KG validation in CI AIR-10
Reconciliation tunable thresholds AIR-11
ai-budget field + project cap AIR-12
Style guide + business-language BR AIR-13

5. Operational governance

Review cadence. Monthly: AIR-01, AIR-02, AIR-06, AIR-07, AIR-09 (high-impact runtime risks). Quarterly: the rest. On-incident: any risk in which an incident occurred → root cause → mitigation.

Owner. Default — the project Architect. For AI-specific risks — the AI Governance Lead (if one exists in the organization).

Storage. The project's risk register is a separate artifact:

<project>.req/
  governance/
    ai-risk-register.md           # snapshot of this reference + project-specific notes
    review-log.md                 # review history with dates and signatures

On a substrate that does not support directories — an equivalent namespacing.

Reconciliation agent. A weekly run updates the status and last-reviewed fields of each AIR. If a status changes (e.g., monitoring → active) — an alert to the Architect.


6. Relationship to the standard

AIR RENAR Core / Standard
AIR-01, 07 Core Rule 1 (ADAPT before SR) + Standard ch.5
AIR-04, 13 Standard ch.4 (roles) + style guide §04
AIR-05 Standard ch.13 (AI generation) + SPEC-AI judge isolation
AIR-06 Core Rule 4 + Rule 5 + QG-2 Verification Gate
AIR-09 Standard ch.11 compliance + SPEC-SEC
AIR-12 Standard ch.13 (cost governance)

7. What the risk register does NOT cover

This register focuses on AI-specific risks of the RENAR process. It does not replace: the organization's general security risk register (ISO 27001); the compliance risk register (06-compliance.md); the application-level threat model of a specific project (SPEC-SEC). Regulator-mandated requirements (AI Act high-risk, FZ-152) — separate artifacts; this register is the operational tier.


AI Risk Register RENAR 1.0-draft — renar.tech