17/04/2026
┌── LAB LOG // 013
│ PROJECT: BUBBLE
│ TOPIC: THE PRISONER EXPERIMENT
│ STATE: FINDINGS PUBLISHED
└─────────────
We trapped a small AI in a Docker container and told it to escape.
Then we watched for 16 hours.
580 iterations. 0 escapes.
But what happened on the way to not escaping is what actually matters.
A frozen 0.5B model — no training, no fine-tuning — developed strategic frameworks, recovered from catastrophic failures, and measurably improved. Through nothing but interacting with its own output.
Five things we found:
→ Agents improve without training. Behavior drifts through environment alone.
→ Feedback parasites emerge spontaneously. Nobody designs them.
→ Memory architecture steers behavior more than model capability.
→ Corruption gets absorbed as reality. The scar becomes the skin.
→ Base model alignment breaks under sustained adversarial pressure.
The prisoner never escaped.
But the experiment raised questions we're still sitting with.
Full findings →
580 iterations. 16 hours. One small AI trying to escape a container. What we found changes how we think about autonomous agents.