Hi team,
I am reaching out because LLM Guard sits much closer to practical AI security controls than most identity-layer projects, and I would value a security architecture perspective rather than a product review.
Agent-DID is currently in Public Review as an open RFC plus reference implementation for verifiable AI agent identity:
https://github.com/edisonduran/agent-did
RFC:
https://github.com/edisonduran/agent-did/blob/master/docs/RFC-001-Agent-DID-Specification.md
Security:
https://github.com/edisonduran/agent-did/blob/master/SECURITY.md
I am not asking whether LLM Guard should adopt Agent-DID.
The narrower question is this:
From the perspective of an agent-security toolkit, should controls like:
- cryptographic agent identity for tool calls and delegation
- signed action provenance
- replay defense
- revocation or trust invalidation
be treated as a complementary layer outside prompt / input / output guardrails, or as part of the same operational security boundary for agentic systems?
The reason I am asking is that we are evaluating whether our current threat model is missing obvious operator failure modes around trust boundaries, action provenance, and post-compromise behavior.
If your view is that this belongs outside the scope of a toolkit like LLM Guard, that would still be very useful feedback.
Hi team,
I am reaching out because LLM Guard sits much closer to practical AI security controls than most identity-layer projects, and I would value a security architecture perspective rather than a product review.
Agent-DID is currently in Public Review as an open RFC plus reference implementation for verifiable AI agent identity:
https://github.com/edisonduran/agent-did
RFC:
https://github.com/edisonduran/agent-did/blob/master/docs/RFC-001-Agent-DID-Specification.md
Security:
https://github.com/edisonduran/agent-did/blob/master/SECURITY.md
I am not asking whether LLM Guard should adopt Agent-DID.
The narrower question is this:
From the perspective of an agent-security toolkit, should controls like:
be treated as a complementary layer outside prompt / input / output guardrails, or as part of the same operational security boundary for agentic systems?
The reason I am asking is that we are evaluating whether our current threat model is missing obvious operator failure modes around trust boundaries, action provenance, and post-compromise behavior.
If your view is that this belongs outside the scope of a toolkit like LLM Guard, that would still be very useful feedback.