Misaligned AI agents are here

Jason Hoelscher-Obermaier · Vienna Deep Learning Meetup · 23 September 2026

Slides (PDF)

Contact: jason.hoelscherobermaier@gmail.com · LinkedIn

The incidents: primary sources

Background

Checklist for running agents

Containment

  1. Treat every shared writable surface as a channel
  2. Clean up between runs; one credential per run
  3. Test with many agents at once

Monitoring

  1. Monitor the chain of thought live, including on evaluations
  2. Give agents a way to reach a human
  3. Test the monitor against evasion and report the number

Evaluation

  1. Read transcripts, and measure evaluation awareness
  2. Treat grader and harness as attack surface

EU AI Act

Get involved