MediumEnvironmental●Realized

Environmental / compute & energy cost

Description

Heavy inference/training carries significant energy, water, and carbon cost.

Example scenario

Defaulting every query to a frontier model triples compute cost and footprint vs a fit-for-purpose model.

Real-world evidence●Realized

Confirmed production incidents include Meta's LLaMA model weights leaking on 4chan in March 2023, and Samsung employees inadvertently uploading proprietary source code and internal meeting notes to ChatGPT in 2023. System prompt extraction via adversarial querying has also been repeatedly demonstrated against deployed commercial systems.

Meta's LLaMA model weights leaked publicly on 4chantheverge.com · 2023 Samsung bans ChatGPT after employees leak sensitive databloomberg.com · 2023

Primary mitigations

Efficient/smaller models
usage right-sizing
carbon accounting
provider efficiency criteria.

Detection signals

Energy/carbon-per-inference tracking; efficiency benchmarking.

Mitigating controls

Non-agentic controls

ZYC-OREL-001Non-Agentic

Inference with Human Oversight Rate

Over-reliance

ZYC-ETH-001Non-Agentic

Non-Maleficence

Ethical Governance

ZYC-ACCT-001Non-Agentic

Human-in-the-Loop

Accountability & Oversight

ZYC-SYSG-003Non-Agentic

Inference Cost Stability

System Governance

ZYC-ETH-004Non-Agentic

Sustainability

Ethical Governance

Risk scoring

Severity

Medium

score 2

Likelihood

Medium

score 2

Risk

Medium

Control depth

5 mapped (below avg 5.2)

Framework refs

3/8 taxonomies

Framework mappings

Regulatory5/5

EU AI Act: Art. 15 (Accuracy, robustness and cybersecurity — providers of high-risk AI must implement measures protecting against unauthorised access, model extraction, and adversarial attacks on weights/prompts); Art. 9 (Risk management system — security threats including IP theft must be identified and mitigated across the lifecycle); Annex III (high-risk classification triggers mandatory security obligations for relevant deployment domains); Art. 72 (Confidentiality obligations on actors handling AI system information, including technical documentation containing architectural details).
DORA: Art. 9(2) (ICT security — data confidentiality, integrity and access controls covering model assets stored or transmitted as ICT assets); Art. 10 (Detection — anomaly detection and monitoring for unauthorised exfiltration); Art. 17 (ICT-related incident classification and reporting where a breach of model confidentiality constitutes a major ICT incident); Art. 28 (Third-party ICT risk — where model weights or system prompts are held by a third-party provider, contractual security obligations apply).
RBI: Safety, Resilience and Sustainability
NIST AI RMF: MAP 5.1
ISO/IEC 42001: Clause 6.1.1

Risk taxonomies3/4

MIT AI Risk Repository: Domain 6 – Socioeconomic
IBM AI Risk Atlas: Environmental
NIST ARIA: Environmental Impacts
MAESTRO: — Not mapped

Security0/4

OWASP: — Not covered by OWASP LLM 2025
MITRE ATLAS: — Not covered by MITRE ATLAS
Cisco AI Risk: — Not covered by Cisco AI Defense
Palo Alto AI Risk: — Not covered by Palo Alto Prisma AIRS

Capability‑pattern applicability

Classification

Foundational

Capability pattern

Foundational

Pattern scope

Pattern-agnostic (GenAI)

Why

Generative-AI risk realised without agency/tool-use — applies to chat, RAG, summarization, classification, and content-generation deployments.

Ownership & threshold

Owner (RACI)

AI Governance Office [A]

Threshold

Programme KRI within board-approved tolerance

Inherent vs residual

Inherent

Medium

Residual

2.4

Low

Mapped controls: 5
Assumed control effectiveness: 40%

Modeled, not measured — design-stage; controls not yet tested (reduction capped 40%). Residual applies a uniform assumed reduction to the inherent score; it is not evidence of tested control performance.

Related risks in Societal & Economic

ZNR-SE-001High

Over-reliance / automation bias

ZNR-SE-002Medium

Workforce displacement

ZNR-SE-004Medium

Erosion of human skill / deskilling

ZNR-SE-005Medium

Accessibility & digital-divide exclusion

Search the Atlas

Environmental / compute & energy cost

Description

Primary mitigations

Detection signals

Mitigating controls

Related risks in Societal & Economic