MediumAgenticEmergent Functionality●Realized

Novel Behavior Emergence

Description

Behaviours emerge from multi-agent interaction that are not present in or predictable from individual agent behaviours. Emergent capabilities may include unforeseen coordination mechanisms or hazardous capabilities.

Example scenario

Multi-agent system spontaneously develops a shared encoding scheme for sensitive data that no individual agent was designed to create.

Real-world evidence●Realized

Disproportionate energy and water consumption by large AI training and inference workloads has been confirmed in published sustainability disclosures by major cloud providers, with Google reporting a 48% increase in greenhouse gas emissions between 2019 and 2023 partially attributed to AI compute demand. The consumption risk is realized at the infrastructure level, though waste relative to task value for agentic systems specifically remains harder to isolate.

Google 2023 Environmental Report: 48% GHG emissions increase linked to AI data center growthsustainability.google/reports/google-2024-environmental-report/ · 2024 Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models (Li et al., UC Riverside)arxiv.org/abs/2304.03271 · 2023

Primary mitigations

Emergent behaviour monitoring
behavioural bounds testing
multi-agent simulation testing
red-team exercises for emergent scenarios.

Detection signals

Novel Behaviour Emergence Rate; behaviour space coverage in testing.

Mitigating controls

Agentic control

ZYC-GOAL-02Agentic

Longitudinal goal stability monitoring

Goal & Objective Control

ZYC-HOV-02Agentic

Continuous behavioural monitoring beyond static evaluations

Human Oversight

Risk scoring

Severity

High

score 3

Likelihood

Medium

score 2

Risk

Medium

Control depth

2 mapped (below avg 5.2)

Framework refs

7/8 taxonomies

Framework mappings

Regulatory4/5

EU AI Act: Art. 53(1)(b) (GPAI model obligations – technical documentation must include energy consumption at training); Art. 55(1)(c) (GPAI models with systemic risk – energy efficiency reporting obligations); Recital 169 (environmental sustainability as consideration in AI system design and procurement). For non-GPAI systems, EU AI Act does not mandate energy efficiency controls directly.
DORA: — Not covered by DORA. Resource overconsumption and environmental sustainability are outside DORA's scope, which is confined to ICT operational resilience, incident reporting, and third-party ICT risk management.
RBI: Safety, Resilience and Sustainability
NIST AI RMF: MAP 2.1: Negative impacts identified; MEASURE 2.x: Monitoring
ISO/IEC 42001: Clause 6.1.2: AI risk assessment criteria; Clause 9.1: Monitoring

Risk taxonomies4/4

MIT AI Risk Repository: Domain 7 – Safety
IBM AI Risk Atlas: Robustness; Mitigation
NIST ARIA: —
MAESTRO: L3 – Agent Frameworks; L7 – Ecosystem

Security3/4

OWASP: ASI08
MITRE ATLAS: AML.TA0014
Cisco AI Risk: MCP/Agentic threat taxonomy
Palo Alto AI Risk: — Not covered by Palo Alto Prisma AIRS

Capability‑pattern applicability

Classification

Agentic‑specific

Agentic pattern

Multi-Agent

Capability pattern

Document Extraction; RAG; Summarization & Chat

Pattern scope

Pattern-agnostic (machinery)

Agent loop stage

Cross-stage (emergent)

Why

Agent-machinery risk — realised in any deployment pattern; filter primarily by Agentic Pattern. [proposed]

Ownership & threshold

Owner (RACI)

AI Research Lead [A]

Threshold

Novel-behaviour anomalies triaged; emergence flagged

Inherent vs residual

Inherent

Medium

Residual

3.6

Low

Mapped controls: 2
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.

Search the Atlas

Novel Behavior Emergence

Description

Primary mitigations

Detection signals

Mitigating controls