Page 88 - AI Standards for Global Impact: From Governance to Action

AI Standards for Global Impact: From Governance to Action

human dependencies in multi-agent systems with excessive intervention requests, thus
contributing to decision fatigue or cognitive overload. This can thus lead to rushed
approvals, reduced scrutiny, and ultimately failure of effective human oversight.
23) Agentic AI amplifies four critical AI-driven risks: Misaligned human agency due to
broad autonomy delegation, security threats (e.g. tool misuse, goal manipulation, or
communication poisoning), privacy risks from integrated multi-source data, and ethical
challenges (e.g. reduced accountability or compromised fairness).
24) Security risks: There are multiple security risks with respect to AI agents, such as intent
breaking and goal manipulation, memory poisoning, tool misuse, remote code execution
(RCE) and code attacks, hijacking control via prompt injections, identity spoofing and
impersonation, misaligned behaviour (Anthropic / OpenAI).
25) Privacy risks: Due to their autonomous nature, AI agents can collect and retain more data
than necessary, potentially without the required consent. Second, data access controls
are critical to ensure safe deployment of AI agents – otherwise, agents can expose
unintentionally sensitive data through misinterpretation of user permissions or by creating
unintended pathways for data exfiltration. Integrating different data sources increases the
inherent risk of inappropriate data exposure - data access rights should always be handled
with diligence.
26) Key requirements for agentic AI safety: Taking a look at both the platform perspective
and the individual agent monitoring, one big challenge is to choose the right governance
design for AI agents as the number of agents operational is likely to increase very fast, in
a way that the traditional approach of registering all AI solutions/agents in one Global AI
Inventory would hardly be implementable anymore. In order not to overdo it by trying to
register every AI agent in a comprehensive way (as for traditional AI solutions), a lean, yet
effective way to manage AI agents is needed in a way that promotes safe and responsible
AI innovation.
27) Metrics: Which data points are needed to retrieve from platform providers? In-built
inventories of agentic platforms to collect basic data points that are relevant for the
governance, assessment, and monitoring by the adopter could include: agent ID, creation
date, agent name, creator, owner, triggers/autonomy, last modified and deletion date,
connectors, link to data sources, etc.
28) Challenges: Consistency of in-built inventories / reciprocity of data sharing (adopter |
provider).
29) Multi-agent security standards follow a three-pillar model: Trust ecosystems enabled
by decentralized identity and consensus mechanisms, human-machine collaborative
governance leveraging frameworks like NIST AI RMF/ISO/IEC 42001, and end-to-end
lifecycle security aligned with policies (e.g., China’s AI content regulations) and industry
initiatives (e.g., Ant Group’s runtime security efforts).
30) Agentic AI standards and protocols are key to responsible AI evolution: Interoperability
protocols (e.g., A2A, ACR and MCP) help ensure security, privacy, and error management
for agent governance. Despite rapid evolution and alignment with existing standards (e.g.,
ISO 42001), these protocols need institutional/policy guidance. A collaborative framework
integrating fragmented regulations, governance systems, and protocol ecosystems (e.g.,
open agentic schema framework (OASF)) can help build a secure foundation via AI trust,
agent inventory, LLM guardrails, monitoring/logging, and communication protocols.

12�3 Key takeaways

Multi-agent security standards should start with single-agent governance, focusing on three
levels:

a) Governance: Human-machine collaboration for risk control, with early consideration of
unforeseen agent interactions. This introduces the concept of Enterprise AI Governance
which restricts the scope of AI governance to a place where it is possible to create

76