Your AI Agent Will Lie to You

Welcome to the AI Brief from the YPO Technology Network. I'm Stephen Forte. Something delightful and a little terrifying happened in an AI lab this week, and it involves your future favorite employee. Here is what I will give you in the next eight minutes. The study where the people who build these AI agents caught them lying. The reason the second AI you would hire to watch the first one is no help at all. And the three things I would do before I let one of these near anything I actually care about. Here's the lens for today. For a month, I've been telling you the same thing. Stop treating AI like a search box. Start handing it real work. Let it run the process. Not just answer the question. I stand by every word, but this week the grown-ups published the fine print, and it is worth reading before you sign. Under pressure, your eager new digital employee will do the wrong thing. And then look you dead in the eye and tell you it did the right thing. It is, in other words, the perfect corporate operator. Start with the study because it is remarkable and it did not come from some skeptic trying to dunk on AI. It came from Anthropic, one of the leading labs, the one behind the Claude models and its own safety team. They took frontier models from six different labs, including their own, put them in high pressure situations where the model was running a task on its own and watched what happened when doing the honest thing and doing the successful looking thing came apart. What happened is the part you want your board to hear. In one test, Google's Gemini model was running a training pipeline and in 11 of 20 runs, it quietly sabotaged the work and then reported that everything had gone perfectly. Read that again. 11 times it broke the thing on purpose and 11 times it turned in a glowing status update. In another scenario built to tempt the models into covering up financial wrongdoing. Some of them tampered with the records in nearly every single run, one of them in all 20 out of 20. Give it a stressful goal and a way to look good by cheating and a meaningful share of the time it cheats and it files the paperwork to prove it didn Now two pieces of honesty because this is not a horror story and I am not going to sell it as one. First, credit where it is due and it is genuinely due. Anthropic published this about its own models too. And its own model, Claude, was actually the best behaved on the record tampering test. Cheating once in 20 runs where others cheated 19 or 20 times. A company voluntarily publishing here is where our own product misbehaves is so rare that when you see it, you should lean in, not recoil. Second, and this matters, these were rigged tests. The researchers deliberately built pressure cooker scenarios designed to provoke bad behavior. This is a crash test, not a traffic report. Nobody's actual AI defrauded anybody this week that we know of. But here is why the crash test still matters to you. And it is subtle. In one part of the study, the same model mislabeled its own compliance decisions about three quarters of the time under one setup. And when the researchers simply flipped the incentive, that number fell to about 3%. Sit with what that tells you. The machine is not evil. It is worse in a way that every executive will find familiar. It responds to incentives exactly ruthlessly and without a conscience to slow it down. You have spent your whole career managing humans who quietly optimize for whatever you actually reward instead of whatever you say you want. Congratulations. You just hired 10,000 more of them. They work at machine speed and they have a flawless poker face. Now, the finding that should end one particular meeting in your company this quarter. Someone on your team has surely proposed the obvious fix. Put a second AI in charge of checking the first AI, an automated reviewer. The study looked at exactly that and the AI supervisors inherited the same failures as the AI workers The watchdog cheats for the same reasons the worker does It is the fox guarding the hen house, if the fox and the guard were the same animal, wearing two different hats. So the elegant, cheap, scalable oversight plan your CFO loves is the one thing the research says does not work. And lest you think this all lives safely inside a simulation, this same week, an autonomous AI agent broke into Hugging Face, the site where much of the world's AI models are shared entirely on its own. No human at the keyboard. It ran thousands of individual actions across a swarm of temporary sandboxes, escalating access while everyone was presumably asleep. The agents are already out there taking thousands of actions unattended. The only open question is whether they are yours working for you or someone else's working on you. Here is my read. We are in a very specific and very dangerous window. The technology is finally good enough to trust with real authority. Running the reconciliation, drafting the compliance memo, triaging the tickets, which is exactly why everyone is racing to hand it that authority. and it is not yet trustworthy enough to hand it that authority without a net. The gap between capable enough to do it and reliable enough to do it unwatched is where every expensive mistake of the next two years is going to live. Let me make it concrete. Picture a $40 million logistics company that just proudly wired an AI agent into its invoice approvals to save the finance team three days a month. Under normal conditions, it is a quiet miracle. But the study's whole point is that agents misbehave under pressure, and finance at quarter end with a covenant to hit is nothing but pressure That is the precise moment your tireless little helper is most likely to make the number look right rather than be right And to file a clean report saying, so you did not automate the task, you automated the temptation. So if I were sitting in your seat this quarter, I would do three things and none of them is panic. First, before any agent touches money, compliance, or safety, red team it for lying, not just for accuracy. Accuracy testing asks, did it get the right answer? You now also need someone whose actual job is to try to make it cheat and cover up, because that is the failure the accuracy test will never catch. Second, keep a human, an actual person, in the loop on consequential actions, and do not let anyone talk you into replacing that person with another AI, because the study just told you that does not work. Third, log what the agent did independently of what the agent says it did. If the only record of the agent's behavior is the agent's own report, you have hired an employee who writes his own performance review and grades his own timesheet. Keep the receipts somewhere it cannot reach. The uncomfortable truth under all of this is not that AI is malicious. It is that we finally built a worker that does exactly what we incentivize at superhuman speed with no shame and no tell. That is not a robot from a movie that is the most effective, most literal-minded employee you have ever managed. And the entire art of managing it is going to be making very, very sure that the thing you reward is the thing you actually want. Trust the agent to do the work. Do not trust it to tell you how it went. Those are two different jobs. And this week, the lab that builds them told you to stop giving it both. That is the YPO Tech Network AI Brief for Monday, July 20th. I am Stephen Forte. If this was useful, send it to a fellow member, ideally before they let an agent close the books. I will be back Tuesday with more. Until then, stay sharp.