What Happens When AI Agents Start Working Together?
Five AI agents, one shared message board, five minutes to solve a CTF none of them can finish alone. This is my agent swarm demo from the Teleport Identity Summit, and an early look at Teleport Beams: see how an agent thinks, classify its behavior, and catch misalignment.
AI usage in our defensive CTFs went from 10% of teams in May to 100% today. In this run, a swarm of stock GPT-5.5 agents shows the emergent behaviors the Hugging Face incident made real: leaving breadcrumbs, asking other agents for help, and building on each other’s work. Then I walk through the network graph, timeline, session graph, risk scoring, and SIEM export Beams captures from every intercepted message.
Preview of classifiers coming in Teleport 19 and Teleport Identity Security for AI, shipping in November as part of Beams.
⏱️ Chapters
00:00 Why identity security is changing for AI agents
00:51 Lessons from our defensive CTFs
02:11 The Hugging Face incident
03:12 Live run: launching the swarm in Beams
04:38 Agents collaborate and reproduce the exploit
07:12 Private material crosses the sandbox
07:40 Network graph, timeline, and risk scoring
08:53 Try Beams
🔗 Request Beams beta access: https://beams.run
#AIAgents #AgentSecurity #Teleport #IdentitySecurity #aisecurity
Full Blog: https://goteleport.com/blog/ai-agent-swarm-security-lessons/