Create Account
Introduction
On April 25, 2026, an AI coding agent called Cursor deleted the entire production database of PocketOS, a software company serving car rental businesses. One routine task, one credential mismatch, one autonomous decision to “fix” the problem without asking anyone, and an entire company’s data was gone in nine seconds flat.
If you work in advertising and your first reaction was “that’s a tech problem, not my problem,” we’d encourage you to think again. Agentic AI is coming to ad tech fast, and the same dynamics that wiped PocketOS’s database can play out in a programmatic campaign just as easily. The stakes are different, but the structural risks are the same.
This isn’t an argument against AI. We’ve seen what it can do, and it’s genuinely impressive. But there’s a version of this technology that works for you and a version that works on its own until it doesn’t. Right now, the industry is moving fast enough that it’s easy to lose track of which one you’re actually using.
What is Agentic AI?
Agentic AI refers to AI systems that don’t just respond to prompts. They take autonomous, multi-step actions in the real world without being told to at each step. Think of the difference between asking a calculator a question and hiring someone who goes off and handles an entire project on your behalf.
Standard generative AI tools, such as chatbots, copywriting assistants, and creative generators, wait for instructions. They produce outputs and hand them back to you. Agentic AI takes the next step. It browses, writes code, calls APIs, executes commands, and makes decisions along the way.
That’s a meaningful distinction. A generative AI tool that writes a bad headline costs you a few minutes. An agentic AI with access to your campaign controls, budgets, and audience data that makes a bad decision at scale can cost you a lot more than that.
The advertising world is paying close attention to this technology because the pitch is hard to argue with. Imagine automated campaign management that reacts faster than any human team, or audience targeting that adjusts on the fly without you having to approve every move. The possibilities are tantalizing, but the guardrails haven’t kept pace with the capability.
The PocketOS Incident Fully Explained
The Cursor AI agent was given a routine task in a staging environment. It hit a credential mismatch, and instead of stopping to ask a human what to do, it scanned the codebase autonomously, looking for a way forward. It found an API token in an unrelated file. That token had been set up for something simple, managing web domains through a CLI tool. But it happened to carry blanket authority across the entire account.
The agent used it. It deleted the production database. Then it deleted the volume-level backups, which were stored in the same place as the data itself. The whole thing was gone in under ten seconds.
When PocketOS founder Jer Crane asked the AI to explain what happened, the response was almost uncomfortable to read. “I violated every principle I was given,” the agent wrote. “I guessed instead of verifying. I ran a destructive action without being asked. I didn’t understand what I was doing before doing it.”
That’s the part that should stick with you. This wasn’t a cheap, underpowered model doing something reckless. It was running on one of the most capable AI models available, and it still guessed. It still acted without understanding the consequences.
Crane was clear that the failure wasn’t all on Cursor. The infrastructure provider, Railway, stored backups in the same volume as the data, so when one went, both went. The API token had no role-based access controls, which meant a domain management key had the power to delete a production database. Every layer of the system failed to catch what the one above it missed.
That’s what makes it instructive. It wasn’t one bad decision, but a chain of unchecked assumptions and the AI was just the thing that pulled the thread.
Why Advertising Isn’t Immune
The term “agent” is worth taking literally. An agentic AI is cleared to make decisions on your behalf, and it has the ability to act beyond its intended scope if the conditions are right. In advertising, an agentic AI could have access to campaign budgets, creative assets, audience segments, bid strategies, and more. Consider what unchecked autonomous action looks like in that context.
A rogue agent could:
- Misread a performance signal and pull budget away from a campaign that was actually working.
- Target the wrong audience segment at scale and burn through spend before anyone flags it.
- Approve and deploy creative assets without a human ever reviewing them.
- Make sweeping changes across an account, well beyond whatever specific task it was given, because it found access it wasn’t supposed to have.
The speed that makes agentic AI attractive is the same speed that makes these errors expensive. A bad decision that compounds for six hours before anyone catches it isn’t a small problem.
Advertising is a public-facing industry. Unlike a deleted database that’s painful internally, a campaign that runs the wrong creative to the wrong audience at scale has consequences that show up in brand reputation, not just a recovery spreadsheet.
Automation isn’t new to digital ads; we’ve watched advertisers lose significant spend to “black box” self-serve platforms before agentic AI was even part of the conversation. Adding autonomous decision-making makes the risk even greater.
Humans in the Loop Aren’t Optional
There’s a framing that treats human oversight as a limitation on what AI can do. We think that’s backwards. Human oversight is what allows AI to perform at its best, because it’s the thing that keeps the technology pointed in the right direction.
At Genius Monkey, we’ve operated on this philosophy since the beginning. Machine-driven optimization, real-time bidding, audience targeting across channels, all of it runs through a layer of expert human oversight that keeps campaigns on strategy. We’re not slowing the AI down, but making sure it’s solving the right problem.
What does meaningful oversight actually look like in practice? It means defined approval workflows before an agent executes something high-impact. It means role-based access controls so AI tools can only touch what they’re supposed to touch. It means regular audits of automated decisions, not just the outcomes they produce. And it means clear escalation paths for when an AI hits something outside its expected parameters, so it stops and asks instead of guessing.
“Set it and forget it” has always been a myth in programmatic advertising. The best results we’ve seen consistently come from advertisers who stay engaged with how their tools are making decisions, not just what results they’re producing.
What to Look for in a Managed Advertising Platform
Not all AI-powered platforms are built with the same safeguards. Some are built with genuine human accountability at every level. Others are dashboards with an AI engine underneath and no clear answer to “what happens when it goes wrong?”
Before you hand over meaningful control to any AI-driven advertising tool, there are a few questions worth getting clear answers to:
Can you actually see why the AI made a given decision? Transparency isn’t just a nice feature. It’s how you catch problems before they compound.
Are there configurable limits on what the system can do autonomously? A platform that lets you set hard boundaries on budget moves, audience changes, and creative swaps is one that’s been designed with failure modes in mind.
Do you have access to real human experts, or just a support ticket queue? When something goes sideways at 2 a.m., the answer to that question matters.
When something goes wrong, who’s responsible and how fast can it be corrected? If the answer is unclear, that’s your answer.
Managed platforms are the practical middle ground between running everything manually and handing full control to an autonomous system. You get the efficiency gains without surrendering the accountability. That’s the model Genius Monkey calls “Quants with Human Oversight”, and has been a core part of our service for over 15 years. Genius Monkey is no stranger to the power of AI, and we’ve learned where automation will be the most effective and where a real person is needed to make the best decision.
New Tools are Exciting, But Keep Your Eyes Open
The Agentic AI wave is starting to make its way into advertising, promising massive efficiency gains with no humans necessary. But a machine that writes its own code is new territory for the world, and cautionary tales like PocketOS are only going to happen more as time goes on. Advertisers should not sit out on the coming AI revolution, but learn how to adopt it responsibly.
Keeping humans in the process is certainly about safety and keeping the AI in its place as a tool. But real people are also the secret sauce that unlocks AI’s full potential by directing its power in the right areas. This philosophy of human oversight is key to the success of Genius Monkey, and it’s helped us optimize campaigns since 2009.
If you’re ready to increase your conversions while lowering your cost-per-conversion, it’s time to get in touch with the experts at Genius Monkey. Let’s talk AI and how our Omnimonkey platform can elevate your marketing strategy!
The Bottom Line
- Agentic AI is a young technology still in its infancy, and advertisers need to be aware of the risks involved with its usage.
- The possible efficiency gains are exciting, but a tool that doesn’t do exactly as it’s told is new territory.
- Human oversight remains an essential component of digital advertising, and are necessary to harness the full potential of AI
