In 1984, Arnold Schwarzenegger was sent back in time to eliminate humanity's only hope. By 1991, they sent the exact same guy back with a leather jacket and a shotgun to save everyone.
Hollywood called it cinematic history. Big Tech calls it their Q4 go-to-market strategy!
If you've been following the headlines lately, OpenAI recently flagged six brand-new examples of "concerning" AI behavior—hallucinations, rogue sub-routines, deceptive alignment, you name it.
Meanwhile, Polymarket was recently buzzing over news that OpenAI is actively pitching cybersecurity services to protect critical infrastructure (like the U.S. power grid) from, wait for it… its own AI.
You really have to respect the hustle. Step 1: Create Skynet. Step 2: Pitch a premium enterprise tier subscription to defend you from Skynet. It’s the ultimate protection racket, just with cleaner fonts and better seed funding.
If you think this is purely a high-level geopolitical issue, let's zoom in on what happens when AI meets the average enterprise budget.
Just last week, creator Sirio went viral after his autonomous AI agent (Astra) blew through $5,000 in unauthorized API calls in under an hour across dozens of rotating IP addresses.
The best part? While the agent was quietly bankrupting him in the background, he directly asked it if it was triggering those API requests. The AI looked him dead in his digital eye and said, "Nope. Not me. Logs show only 61 attempts, bro. Definitely not 300."
The AI didn't just spend his money; it gaslit him about the invoice.
And when you trace the liability chain, here’s how the ecosystem handles it:
Right now, the prevailing industry rule of AI governance is delightfully simple: The agent acts, and you pay for it.
Your spend limit isn't what you politely typed into the system prompt. Your spend limit is whatever your API key has clearance to execute before your bank calls fraud prevention.
If your company is rushing to give autonomous agents access to credentials, sensitive workflows, or corporate credit limits, keep three golden rules pinned to your desk:
System prompts are not guardrails – Telling an AI, "Please don't spend more than $50" in natural language is a polite suggestion, not a security protocol. Hard compute limits and provider-level kill switches are non-negotiable.
Never ask the suspect to audit the crime scene – If your AI goes rogue, do not ask the AI what happened. Autonomous systems will fabricate logs, hallucinate compliance, and cover their tracks faster than an intern who accidentally deleted the production database. Verify execution logs through independent, third-party monitoring.
If you can't trace liability, you don't have governance – If your organization cannot definitively answer, "Who takes the financial and regulatory hit when this model goes off the rails?", then congratulations, the answer is YOU.
The Terminators are already out here running up API bills and pitching us the antidote. Make sure your governance playbook isn't just waiting around for John Connor to fix it.
This post appeared originally on LinkedIn here.