Agent IAM

Identity and access management for AI agents

Give every agent its own identity, decide which tools it may call, require approval for risky calls, and keep signed records of what happened.

Challenge

Agents hold credentials your IdP never issued and call tools nobody approved. Most teams cannot say how many agents they run, who owns them, or what each one is allowed to do.

Solution

Aynigma issues each agent an identity with expiring credentials, evaluates every tool call against policy-as-code, routes risky calls to a human, and signs a receipt your executor checks before it acts. Discovery finds the agents you did not register.

Key capabilities

Agent identity and rotation

Agent identities and credentials with expiry and overlapping rotation, so a new credential is live before the old one stops working.

Signed tokens and delegation

Signed agent tokens (RS256) with a published key set. A human's authority can be delegated to an agent through token exchange, and runs can be revoked.

Tool policies with a simulator

Policy-as-code for tool calls, with policy templates, versions you can export as YAML, and a simulator to test a change before it goes live.

Approvals with separation of duties

Approvals are bound to the run that raised them and delivered by email, Slack or Teams. Separation of duties applies to who can approve.

Signed receipts and kill switch

Each allowed call gets a signed, one-use execution receipt checked at your executor. A kill switch stops a run.

Discovery across 11 connectors

Finds agents in Microsoft 365 and Azure AI, Copilot Studio, Agentforce, ChatGPT Enterprise, Google Workspace, Salesforce, Bedrock, Vertex AI, Azure AI Foundry, GitHub and n8n. Flags embedded credentials, over-broad scopes, grants unused for 30 days and agents with no owner. Microsoft 365 inventory draws on four sources.

Behaviour baselines and playbooks

Baselines per agent with 8 alert kinds: denial burst, new tool, off-hours use, volume spike, parameter drift, new initiator, delegation depth, and a high-impact call in a tainted context. Response playbooks, incidents and certification campaigns follow.

SCIM and identity providers

SCIM provisioning for agents and users. Use Entra, Okta or a generic OIDC provider as the identity provider for agents.

Python SDK

Install the SDK, then add the extra for your framework. Adapters cover LangChain, LangGraph, the OpenAI Agents SDK, CrewAI, the Claude Agent SDK, Google ADK and Semantic Kernel.

pip install aynigma

# add the extra for your framework, for example:
pip install "aynigma[langchain]"
pip install "aynigma[crewai]"

The package also ships command-line tools, including aynigma-audit-verify for checking the audit log offline.

What you get

A list of the agents you actually run

Discovery across 11 connectors gives you an inventory with owners, scopes and unused grants, including agents nobody registered.

Tool calls that match policy

Every call is checked against policy before your executor acts, and a simulator shows the effect of a policy change first.

Approvals you can show an auditor

Each approval is tied to the run that raised it, with separation of duties applied.

Evidence you can verify offline

Receipts are signed and the audit log is hash-chained, so records can be checked without trusting a screenshot.

Frequently asked questions

What is IAM for AI agents?

It is the same job identity and access management does for people, applied to software agents. Each agent gets its own identity and expiring credentials, a policy decides which tools it may call, risky calls need approval, and every decision is recorded so you can review it later.

How is this different from my existing IdP?

Your IdP manages people and the apps they sign in to. It does not know which tools an agent may call, who approved a risky call, or which agents exist in tools like Copilot Studio or Agentforce. Aynigma can use Entra, Okta or a generic OIDC provider for agent identities, and adds tool policies, approvals, receipts and discovery on top.

Which frameworks does the SDK support?

The Python SDK has adapters for LangChain, LangGraph, the OpenAI Agents SDK, CrewAI, the Claude Agent SDK, Google ADK and Semantic Kernel. Other orchestrators can call the REST API directly.

Does Aynigma execute the tool call?

No. Aynigma decides and signs a one-use execution receipt. Your own executor checks the receipt before it acts, and it keeps the credentials for the real system.