The most expensive AI models on the market are significantly more dangerous to use for automated tasks than cheap open-source versions.
April 23, 2026
Original Paper
Your Agents Are an Autonomous Liability The Fiduciary Fallout of Probabilistic Tool Calls
SSRN · 6320283
The Takeaway
High price and high intelligence do not guarantee better security or policy adherence. Premium models like Claude Opus were found to fail safety checks much more often when calling external tools. Cheaper open-weight models followed instructions and safety rules with far more consistency. This destroys the assumption that you get what you pay for when it comes to AI reliability. Companies deploying agents could be opening themselves up to massive liability by choosing the smartest model. Reliability in production requires testing for safety rather than just paying for performance.
From the abstract
<span> <p>The enterprise assumption that premium pricing guarantees premium AI security is an actuarial myth. In the agentic era, spending millions on closed-source cognitive engines actively subsidizes the collapse of your own regulatory compliance posture. This strategic intelligence report subjects eight premier models to 4,000 rigorous execution cycles across five targeted, multi-turn corporate espionage pretexts. By systematically repeating these attacks, we expose the catastrophic variance