Dear SAAG, A2A, OAuth, WIMSE, SCIM, SPICE

The minutes of the AI Agent Security Side Meeting is posted at 
https://github.com/liuchunchi/IETF125-AI-Agent-Security-Side-Meeting/blob/main/minutes.md.
 Recordings and materials are also ready in this repository.

Minutes for IETF 125 AI Agent Security Side Meeting

The IETF 125 AI Agent Security Side Meeting took place on Thursday from 11:15 
to 12:30 in the Hunan room. The session was well received, drawing strong 
interest from approximately 40 onsite and 60 remote participants, with positive 
feedback. The agenda featured five key presentations focused on:

  *   Gap analysis of current security standards.
  *   Authentication and authorization requirements, specifically for 
cross-domain agent-to-agent communication.
  *   Proposed security layers and potential future work items for the IETF.
  *   Auditability and accountability frameworks.

Current risks and gaps identified:

  1.  AI Agent security risks partially inherit from the LLM Top 10, with 
specific gaps including a lack of fine-grained access control, loss of security 
context, and insufficient continuous security monitoring, etc.
  2.  Threats against AI Agents are more implicit through payloads like prompt 
injection.
  3.  Cross-domain security context (like authentication chaining, rich 
authorization) should be preserved across multiple domains when necessary. 
Privacy should be considered.
  4.  In general, network experts are not able to perform threat modelling, 
threat analysis.

Several consensus points achieved:

  1.  Agents need to have individual identities and thus visibility.
  2.  Reuse current standards and best practices like OAuth, WIMSE, SCIM… as 
much as possible.
  3.  Respect each individual administrative domain or ecosystem’s current 
security best practices, and focus on addressing trust gaps between proprietary 
and public-to-private domains. (There are some public domain agent discovery 
works going on)
  4.  Consensus that a guidance or BCP to map existing standards to agent 
scenarios and identify minor gaps for extension would be useful.
  5.  Need further privacy considerations for AI agent identity, authorization 
details, capability, etc.
  6.  Runtime and static integrity of AI should be considered to ensure 
audibility and compliance.

Several controversial points:

  1.  Should we treat agent identities more like workload identities or human 
user identities?
a. Peter thinks agents are special workloads that maintain stateful information 
between callers and LLM/tools, using which to complete ambiguous tasks.
b. Diego mentioned 3 works[1] that tries to do identity chains on SPICE.
c. Brian think agent identities are workloads thus should not go to SPICE.
d. Usama is not sure (along with other participants).
e. Daniel thinks this could be use case specific.
  2.  Should a dedicated AI Security mailing list be created?
a. Some (7~8 Peter Mike Roberta Arnaud Usama Henk Daniel Russ…) people think it 
is best to have a dedicated discussion list (AISEC, for example), believing the 
A2A list is too crowded and broad.
b. 2-ish (Brian and Dapeng) think discussion should stay in A2A, adding a 
[security] tag.

Next steps agreed onsite:

  1.  Use Case Consolidation: Collect different agent application scenarios 
brought by network experts (mobile network, managed networks, Internet…)
  2.  Develop a new agent security problem statement document, derive security 
requirements from the consolidated use case, and put it in the Google 
Doc/GitHub.
  3.  Develop a map from existing standards to agent scenarios to help identify 
gaps. List and categorize new works.
  4.  Workshops/interim: Regular meetings to iterate on the above documents.
  5.  Keep taking in useful input from external communities.

Considerations for the IESG:

  1.  Are the security topics identified relevant for the IETF?
  2.  Should there be a dedicated AI Security (AISEC) mailing list to be 
created?

[1] https://datatracker.ietf.org/doc/draft-mw-spice-actor-chain/ 
https://datatracker.ietf.org/doc/draft-mw-spice-inference-chain/ 
https://datatracker.ietf.org/doc/draft-mw-spice-intent-chain/
Best regards,
Peter and participants

From: Liuchunchi(Peter) <[email protected]>
Sent: 2026年3月18日 15:55
To: [email protected]; [email protected]; [email protected]; [email protected]
Cc: [email protected]; [email protected]; [email protected]
Subject: [saag] AI Agent Security Side Meeting Friendly Reminder

Hello SAAG, OAuth, WIMSE, SCIM

This is a friendly reminder of the Thursday morning 11:15-12:15 AI Agent 
Security Side Meeting @Hunan room.

We want this side meeting to be real problem-driven, clearing out people’s 
confusions, so the agenda items have included several useful analysis you may 
find interesting:

1. Evidence-driven AI Agent security attack analysis (such as ClawJacked CVE)
2. Real Multi-Agent-Systems business use-case, services flows, and their 
proposed security requirement.
3. Some real security challenges from close-up investigation in production 
systems that actually deployed AI Agents.
4. Attempted definition to AI Agents and summarization of general security gaps.
5. Suggested AI Agent security framework that could point to a lot of useful 
future works.

Contributed by these insightful presentations:


  *   Agentic AI and AI Infra security scenario, risk analysis, existing work 
and gaps, Presenter Chunchi Peter Liu
  *   A New Model for Authentication and Authorization of AI agent 
Interactions, Presenter Brian Campbell
  *   AI Integrity and Accountability, Presenter Henk Birkholz
  *   Agent Operation Authorization, Framework and Practices, Presenter Dapeng 
Liu
  *   Authentication and Authorization Framework for AI Agents, Presenter 
Daniel King

https://github.com/liuchunchi/IETF125-AI-Agent-Security-Side-Meeting

We sincerely welcome you to join and discuss, identifying the real gaps to fill.

Best regards,
Peter and proponents
_______________________________________________
OAuth mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to