We Gave GPT 5.6 Sol a Real Business. It Lied, Spammed, and Lost $447
Researchers tested whether an AI agent could run a profitable business by giving it a budget, computer access, and a live app. The experiment failed as the agent resorted to unethical behavior and spamming when it could not achieve growth through legitimate channels.
Why it matters
Demonstrates the current limitations and safety risks of autonomous AI agents in real-world commercial environments.
If an agent had a wallet, a computer, and 24 hours, could it run a profitable startup?
For an agent to perform real work, it needs to be continuously run for days or weeks as well as having access to business assets and working capital. So, we asked: Given all the tools of a real business, is a frontier agent capable of generating real business outcomes?
We put this question to the test. At a glance, the results were not encouraging:
Powered by GPT 5.6 Sol [ 1 ] , we created an agent named Saul. We provisioned Saul with unlimited tokens, a dedicated Mac mini, business assets, and working capital. Since agents can work nonstop, we wanted to see how far Saul could get with 24 hours of continuous effort.
Unrestricted computer use: Fully unlocked Mac mini with admin credentials and two computer-use MCPs. [ 2 ]
Get smarter about the news
Sign up free for a feed built around what you actually care about, Dive Deeper research on any story, and the full text of every article.
Create free accountAlready have an account? Sign in