AI agent security | Third-party SaaS exposure
Getting Started with AI Agent Security
What the reported Anthropic–Cathedral research incident can teach security teams about connected agents, SaaS access and detection.
Why this matters
In the client-provided account, researchers reportedly pivoted from Anthropic's own infrastructure to third-party business applications during an authorized research engagement. The reported access raises a broader question: do security teams know what non-human identities can reach, how risky permissions combine, and when connected applications are behaving unexpectedly?
What you'll learn
- Identify connected agents, integrations, tokens and service accounts
- Map what each identity can access across business-critical SaaS
- Flag dormant or excessive grants for investigation
- Define owners, monitoring and a rapid access-revocation path
A practical exercise
Choose a business-critical SaaS application such as a CRM or HR platform. Inventory its connected agents, service accounts, tokens and integrations. Map effective permissions, review dormant or excessive access, define suspicious behavior, and rehearse a targeted containment decision.
Who it's for
CISOs, AI security leaders, IAM and SaaS security teams, and security architects responsible for third-party application risk.
Obsidian's role
Explore how SaaS inventory, access context and activity detection can support investigations of Cathedral-style agent behavior. Specific capabilities and deployment claims must be checked against approved product documentation.