Teleport has expanded its Identity Security platform with three new capabilities designed to ensure that agent behavior remains within defined boundaries: Beams Session Summaries, Agentic Classifiers, and Risk Scoring.
They give enterprises a foundational harness for identifying and preventing agent misalignment as autonomous agents take on greater responsibility inside production infrastructure.
The announcement follows Teleport’s recent white paper, From Zero Trust to Agent Trust, which argues that zero trust is necessary but insufficient to govern agents operating at scale. The paper extends the three core principles of zero trust into three corresponding principles of agent trust:
- Verify explicitly → Enforce continuously. Agents need a unique, attestable identity and must operate inside a trusted runtime that architecturally enforces their operational, execution, and communication boundaries.
- Use least privileged access → Bound collective autonomy. Individually authorized actions can still be collectively destructive when taken by a swarm of agents acting in parallel. Bounding collective autonomy means actions that are safe individually but risky in aggregate require escalation before they execute.
- Assume breach → Assume misalignment. Agents can drift from their original objective through adversarial manipulation or through unintentional causes, like context shift over time. Enterprises must continuously monitor for that drift and be able to intervene in real time.
Teleport’s new capabilities are delivered through Beams, Teleport’s trusted runtime for agents, working in concert with its Identity Security platform, now extended to address agentic behavior:
- Beams Session Summaries are a summary of an AI agent’s actions: its identity, privileges, tool and API calls, LLM prompts, responses, and reasoning digested into a short human readable summary of what it was doing and what it was thinking. This establishes a behavioral baseline for evaluating agent activity against its declared objective.
- Agentic Classifiers provide policy for humans, agents or groups of agents to be evaluated against company specific criteria, enabling agent behavior to be flagged that is inconsistent with an agent’s declared objective.
- Risk Scoring automatically summarizes SSH, Kubernetes, and database sessions, classifies them by risk level and maps actions to the MITRE ATT&CK framework. Infrastructure and Security teams can now automate or manually search across sessions for specific commands, resources, or behaviors.
These capabilities in concert with Beams lay the foundation for the agent trust principles: they give agents a cryptographic, continuously monitored identity; they make collective and individual risk visible before action is taken; and they give enterprises the tooling to detect and respond to misalignment as it happens, turning “assume misalignment” from a design principle into a running practice.
“The capabilities we’re announcing today are the operational harness for agent trust: they let enterprises see what an agent actually did, classify whether that behavior is expected, and score the risk of what it might do next,” said Ben Arent, Director of Product at Teleport.
“Zero trust assumes the actors inside the architecture are human, bounded, and verifiable at the point of access. Agents break that assumption. They’re not predictable, and a swarm of individually authorized actions can add up to an outcome no one sanctioned.”
