OAIC 2026 / Full talk / 27:02

Teaching Agents to Escape

Name: Teaching Agents to Escape: RL, Command Evasion, and the Limits of the Sandbox
Uploaded: 2026-10-22
Duration: 27 min 2 s
Description: Gaspard Baye examines how offensive agents search for command-policy bypasses, and why bypassing a filter differs from escaping a capability sandbox. Using SandBench, supervised fine-tuning and reinforcement learning, he compares defenses, analyzes reward-design pitfalls, and explores Bayesian attack graphs as a next step.

RL, Command Evasion, and the Limits of the Sandbox

Gaspard Baye, Ph.D. · Valix AI

About the talk

Gaspard Baye examines how offensive agents search for command-policy bypasses, and why bypassing a filter differs from escaping a capability sandbox. Using SandBench, supervised fine-tuning and reinforcement learning, he compares defenses, analyzes reward-design pitfalls, and explores Bayesian attack graphs as a next step.

Recorded at Offensive AI Con 2026. Includes the full presentation and audience Q&A.

Watch on YouTube ↗

The next wave of offensive AI