Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

I ran the same internal AI workshop twice in one week. Same content, same agenda, same Zoom link, just offered at multiple times to cover different schedules. By the second session, half of my material was wrong. Between the two scheduled sessions, something had changed. In the first session we couldn't use the skills feature in Claude. By the second, we could. Nothing dramatic happened.

In Part 1, Your AI Agent Needs to Know Who You Are, we showed how Teleport JWTs give MCP tools a verified identity for every request. This post extends that pattern to AWS, specifically to Amazon Bedrock AgentCore, where the same identity gap exists but requires a different solution stack. You ask an AI agent to list your S3 buckets. The agent calls an MCP tool. The tool reaches out to AWS. However, CloudTrail records the action under something like agentcore-bot, but not your identity.

In Part 1 of this series, we addressed 18 Kubernetes best practices spanning across container hardening, observability, availability, and fault tolerance. Those practices secure the containers that agents run in. But the CNCF AI Technical Community Group's cloud-native agentic standards go further, establishing that securing containers is only the beginning.

In March 2026, attackers from the TeamPCP group compromised Trivy (CVE-2026-33634) — a widely-deployed open-source vulnerability scanner running in thousands of CI/CD pipelines — and turned it into a credential harvester. SSH keys, Kubernetes secrets, cloud tokens — secrets accessible to any pipeline that ran a compromised version — were exposed. The attacker retained access long enough to exfiltrate newly rotated secrets before the window closed.

AI agents querying databases pose well-documented risks. What gets less attention is the fact that PostgreSQL has no native concept of an agent as a distinct actor. This means DBAs are managing access for something that appears in pg_stat_activity like any other role created with CREATE ROLE, with no distinguishing attributes and no indication of who or what initiated the connection. AI agents have no distinct identity when interacting with PostgreSQL.

Cloud service providers preparing for FedRAMP 20x are encountering a fundamentally different authorization model than the one their compliance programs were built around. The traditional FedRAMP path produced lengthy System Security Plans, point-in-time assessments, and human-readable narrative evidence.

When your AI agent calls an MCP tool, that tool has no idea who actually triggered the request. It sees the agent, not you. This post explains why that matters and how to fix it with Teleport JWTs. In part two of this post, we will explain how to extend this to AWS to carry your identity through Amazon Bedrock AgentCore all the way into CloudTrail.