Agents as Toddlers. How we get audit and visibility into risky agent action. | Teleport
Teleport released three new features for its Beams runtime to track what AI agents do, detect unexpected behavior, and score risk levels. The webinar covers how to audit agent sessions, write policies that catch workarounds, and triage activity at scale using MITRE ATT&CK mapping.
- Session Summaries compress multi-hour agent logs into 30-second summaries with operations, security actions, and risk levels
- Agentic Classifiers catch agents attempting policy workarounds through LLM proxy-level audit across Claude, OpenAI, and other models
- Risk Scoring tied to MITRE ATT&CK framework lets small teams triage agent activity without manual review of every session
Teleport released three new features for their Beams runtime that tackle the core problem of running AI agents in production: you need to see what they're doing, judge whether it's appropriate, and assess the risk before something goes wrong. Session Summaries compress three-hour agent sessions into thirty-second summaries showing what operations ran, which security actions triggered, and the overall risk level. Agentic Classifiers watch agent behavior and catch when they try to work around your policies. Risk Scoring maps agent activity against the MITRE ATT&CK framework so teams can actually triage what matters when dozens of agents are running simultaneously.
The technical approach matters here. Session Summaries work at the LLM proxy level, which means they capture activity across different model implementations—Claude Code, OpenCode, Pi, Codex—rather than requiring custom integration for each one. This matters because teams don't want to rewrite monitoring for every new model that ships. Agentic Classifiers let you write policies that specifically target agent workarounds, not just surface-level rule violations. The MITRE ATT&CK mapping gives small teams a framework borrowed from security incident response, letting them apply proven threat categorization to AI behavior.
The webinar pitch hinges on a specific argument: zero trust security—the foundation of modern infrastructure security—stops working once your security subjects become autonomous. A human operator follows rules and makes decisions you can reason about. An agent running in a loop makes thousands of decisions automatically. You can't inspect every one, so you need different tools: the ability to sample sessions quickly, detect behavioral anomalies, and understand which agent actions actually matter for your risk posture.