Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

Who has the edge in AI security when attackers and defenders have the same models? In this red team vs. blue team panel, Teleport's Ben Arent (Director of Product) talks with Rob (Director of Security, Teleport) and Marcus (founder, maker of Nemesis) about how AI is reshaping infrastructure security: frontier models like Mythos, the Hugging Face incident, prompt injection and the "lethal trifecta," alert fatigue in the SOC, and Teleport's "Pressure Washing" project, where 13 engineers spent a quarter hunting vulnerabilities with LLMs and found more high-severity bugs than in 2024 and 2025 combined.

Watch five AI agents tackle a CTF and start coordinating with one another. Learn how to detect risky agent behavior using @goteleport graphs, risk scoring, and custom classifiers. Ben Arent explores lessons from the Hugging Face incident, then demonstrates how agent identity, scoped access, and activity monitoring help secure AI agents running against infrastructure.

70% of organizations give AI more access than a human in the same role would get. Not surprising when it is common for a team to click “always” when an agent asks to be allowed once or always allowed. That means the agent holds that access in perpetuity and turns into agent over-provisioning. Makes you wonder: Does an agent have more access to your organization's stack than it should and more access than the people who deploy them?

In this video, we'll look at a few key updates from Teleport's latest release v18.11.

We asked eight security and product leaders who's accountable when an agent ships to production and breaks something. Nobody said the model. Harish Gaggar named the reason. An agent runs on permissions someone approved and configuration someone set. Ron Reiter drew the line in the same place, accountability sits with whoever decided what the agent could actually do. As agents act across more systems, the accountability trail gets harder to follow. Most teams cannot determine which human granted an agent access.

What part of your profession do you hope AI takes over? What should it never touch? Security practitioners and technology leaders landed in the same place: hand off the tedious work, and keep human judgment on anything that affects blast radius. That's the skills reckoning. As AI absorbs the repetitive work, the expertise that decides what an agent is allowed to reach becomes more valuable.

The next big breach won't be a hacker. It'll be an over-permissioned agent. Someone approved an agent action at some point, for a reason that made sense at the time. But access persisted long after a task was performed. Stale permissions without human intervention could turn into exposure. Listen to perspectives from people who see this problem from different roles.

This started as an employee contest at our all-hands and turned into one of the best things we've ever made. Meet the Teleportopia house band and yes, you might recognize some likenesses to real Teleporters. Don't make AI a silo. Give it an identity.

Pi is a minimal coding agent that lets you choose the features Claude Code and Codex bake in for you. Nothing is forced, nothing is hidden, and you decide what your harness does. However, Pi runs with no safety rails by design, so the environment has to be the boundary. In this video, we'll take a look at Pi, what it is, its features, how to get your own customized setup running, and why it works so well with isolated agentic environments like Beams.

In this video, we'll look at a few key updates from Teleport's latest release v18.8.