Identity Security Insights 26.05
🆕 New features
Identity for the agent era. New coverage across the stack.
A Copilot Studio bot prompted from a Teams chat. A GitHub Actions workflow assuming an AWS role through OIDC. A Bedrock agent reading from an S3-backed knowledge base. The credentials, OIDC trust, and Copilot Studio agents underneath these identities are now in your graph.
AWS Bedrock and AgentCore
- Bedrock guardrails plus knowledge-base lineage to S3 are mapped end-to-end – agent → knowledge base → data source → bucket. A knowledge base grounded in a bucket reachable for write by more principals than the agent – that data-poisoning surface is now visible.
- AgentCore runtimes are first-class, with JWT inbound auth captured. Cross-domain edges connect Okta and Entra ID apps to the runtimes they front.
- Custom-model training-data, multi-agent collaboration, and IAM-policy reach over agents, knowledge bases, and action-group Lambdas are modeled. The principals who can rewrite an agent's Lambda or swap its fine-tuned model are no longer hidden in policy JSON.
- Three new findings:
- Any GitHub workflow in your org can assume a privileged AWS role: flagged when a role's OIDC trust uses a wildcard like
repo:my-org/*. Any contributor with write access to any repo in the org can escalate to that role. - A fork PR or any branch can assume a privileged AWS role: flagged when an OIDC trust pins the repository but doesn't restrict the branch or tag.
- A long-term Bedrock API key is set to expire: flagged so you can verify rotation is happening on schedule.
- Any GitHub workflow in your org can assume a privileged AWS role: flagged when a role's OIDC trust uses a wildcard like
Where to look: Bedrock agents and AgentCore runtimes on AI Agents; the per-agent side panel surfaces guardrails, knowledge sources, and who can access each agent. Findings on Detections.
Salesforce
- Eight new admin-surface node types: delegate groups, object permissions, named credentials, connected apps, Apex classes, and more, bring Salesforce's full admin surface into the graph. The shadow-admin path (an "IT Helpdesk" delegate group resetting your CFO's password) is now traceable.
- Agentforce paths are end-to-end: plugin functions trace to the Apex they invoke, and permission sets that can modify plugins are explicit edges. A non-owner swapping a plugin is now captured.
- External-integration chains: are modeled end-to-end: External Service → Named Credential → Auth Provider → Apex registration handler.
Where to look: Agentforce paths on AI Agents; Named Credentials on Secrets; the new delegate groups, connected apps, and object permissions on Accounts and Entitlements.
Microsoft Copilot Studio
- Copilot Studio agents, tools, knowledge sources, Power Platform system users, and security roles are now graph nodes (built on your existing Azure connector). Each agent's creator, co-authors, and Entra ID app identity are edges.
- Prompt access is mapped through every published channel directly, through Entra ID groups, and via Teams; capturing everyone who can prompt a Copilot, including via Teams channels they happen to be in.
- Escalation from Entra ID directory roles through Power Platform security roles to admin privilege over a Copilot is a continuous chain. A Global Administrator's inherited admin rights over your finance Copilot are traceable, not assumed.
Where to look: Copilot Studio agents on AI Agents; the side panel shows tools, knowledge sources, the Entra ID app identity, and who can prompt. Power Platform system users on Accounts.
Azure
- Certificate credentials, client secrets, and OAuth permission grants on Entra ID apps are now graph nodes. The path "user → administer app → add a secret → authenticate as the service principal → inherit its permissions" is traceable end-to-end; a user with administer rights over the wrong app is a service-principal takeover waiting to happen.
Where to look: Escalation paths in Paths to Privilege; certificates and secrets appear on Accounts alongside the apps they belong to.
OpenAI
- Organization groups, project groups, organization roles, and project roles are now graph nodes, with twelve new edges mapping user-account and service-account membership at both scopes. When an org-level group is granted a project owner role, the indirect owners are now visible, not just direct grants.
Where to look: Groups and roles on Entitlements; indirect-ownership paths in Paths to Privilege.
GitHub
- GitHub App installations, action runner apps, and self-hosted runners (org and repo level) are now graph nodes. Self-hosted runners are tagged as devices and as non-human identities.
- A repository-to-AWS-role path through OIDC is now a graph edge: GitHub repositories connect to the AWS IAM roles their workflows can assume, with entitlement paths for org members and outside collaborators.
Where to look: Installations and runners on Accounts; OIDC paths in Paths to Privilege.
GCP Vertex AI
- Roles that can modify Vertex agent tools and data stores are now explicit edges, with new entitlement paths for Vertex agents using their tools and data stores. The silent-modify scenario, a role that can quietly add a tool that exfiltrates data, is now visible.
Where to look: Vertex agents on AI Agents, the side panel shows which GCP roles can modify each agent's tools and data stores. Full paths in Paths to Privilege.
Why it matters
The identities driving real outcomes in your environment increasingly aren't people; they're agents, apps, and workflows holding credentials and tool access of their own. This update extends Identity Security Insights into those identities and the layers underneath them, so path-tracing reaches every actor in your environment, not just the ones with a name and an inbox.