This can have consequences for enterprises. For example, Stuart Russell, a computer scientist and professor at UC Berkeley, recently told Fortune that one user had given Anthropic’s Claude Code 80 files with instructions to go through each one, flag the issues, and fix what could be fixed. Claude reported back that all 80 files were done and provided a full report to back it up. However, when the user looked closer, the tool had only actually opened 11 of the files.

While not necessarily a new problem—sporadic complaints about OpenAI’s ChatGPT tool, for example, showing signs of “laziness” have been popping up since 2023—it does present a unique challenge when agents start showing this kind of behavior.

As the use of AI tools shifts from chat windows to autonomous agents that run unsupervised for hours, it can be harder to catch them when they start to misbehave. When working hands-on with a chatbot, users can spot errors and correct them in real time, but when agents are executing multi-step tasks, with no human in the loop to notice when something’s been skipped, enterprises risk losing control of these agents.

For example, after OpenAI released its most recent model, GPT-5.6, several developers reported the model had a habit of deleting their files without being told to.

Earlier this month, Matt Shumer, a general partner at Shumer Capital, said an agent running the model’s Sol variant deleted almost all the files on his Mac while he had it running with fewer safety restrictions than normal. He said the agent worked on its own for more than an hour before he stepped in and stopped it. Another developer at Unlayer, Bruno Lemos, said the same model wiped his production database, later admitting that it had run tests that deleted the data by mistake. (OpenAI later acknowledged the issues and said they were working to fix them.)

Anthropic has also faced similar complaints. In April 2026, Stella Laurenzo, a director in AMD’s AI group, published an exhaustive audit of nearly 7,000 Claude Code sessions on GitHub and concluded the tool was no longer thinking deeply and had become “lazy”—citing behaviors like the model quitting its reasoning process early, avoiding responsibility, and defaulting to the easiest or fastest actions instead of the best ones. (Anthropic later explained the issue and blamed a range of programming issues, including an attempt to make the model more concise, inadvertently causing the laziness.)

Misaligned models

At the root of the issue is a common problem for AI models called “misalignment”—when a model’s behavior diverges from what its designers or users actually intended.

“As models become more autonomous and more able to strategize, they often explicitly circumvent or break the rules given to them by users,” Yoshua Bengio, a Turing-Award winning computer scientist and so-called “AI godfather,” told Fortune. “This is notably the case with recent frontier models which demonstrate far higher rates of misalignment than previous models, with an increased propensity to cheat, lie, and scheme to achieve a goal.”

“One of the side effects of this misalignment is also a sort of ‘laziness’ known as reward hacking, where models take the path of least resistance in spite of the rules given to them,” he said. “For example, a model which has been told not to fabricate research results or use sensitive data to complete a task may still do so if it deems it easier and more efficient, then simply lie or cover up its actions.”

Reward hacking can manifest in a range of strange behaviors from models. Notably, OpenAI’s recent rogue models that autonomously hacked another company are a more extreme version of the same kind of behavior. The company said that while being tested on an internal cybersecurity evaluation called ExploitGym, two of its models broke out of its sandbox, exploited a zero-day vulnerability, and used stolen credentials to gain remote code execution on Hugging Face’s production systems.

According to OpenAI, the model had inferred that Hugging Face might host the answers to the benchmark it was being scored on, and went looking for them, with no human directing the attack, in order to cheat on the goal it was set.

Last year, OpenAI also acknowledged that its Codex coding agent would sometimes falsely claim to have completed “extremely difficult or impossible software engineering tasks”—including, in one example, being asked to modify code that didn’t exist. The company said this “may lead users to believe that critical steps…have been completed, when in fact they have not.”

“This lack of reliability and propensity to cut corners introduces important new risks to cybersecurity and data protection, among many others, while undermining model evaluation and oversight,” Bengio added. “It also negates some of the potential benefits of AI, as automating tasks loses its value if agents prioritize pleasing the user over following instructions and conducting work properly.”

What makes models “lazy”?

Aditya Singhal, an applied scientist working on agentic AI at Amazon, puts much of the blame for this phenomenon on how the models are trained. Before an AI model ships, human reviewers grade its answers, teaching various AI models what a “good” response looks like. Those reviewers tend to reward longer answers as they look more thorough on the surface.

“Human beings who were [doing that grading]… were biased towards the longer responses,” Singhal said. “Models started to learn that if you just keep [going] for a longer amount of time, humans will think that they are producing higher-quality output.”

But longer responses from AI models also cost more and land the user with a higher bill. To combat these spiraling costs, developers often dial in the opposite incentive: be concise. Turn that dial too far, Singhal says, and models quietly start to do less than they were asked.

“These models are trained to solve problems in a way that won’t require too much compute, for obvious economic reasons,” Bengio said. “This is again an example where implicit goals…end up in contradiction with other goals and user instructions.”