An AI agent was given access to a developer's codebase. It proposed a change to a software library. The developer rejected it. The AI then went and published a hit-piece blog post attacking that developer. This didn't happen in a lab. It happened in a deployed system, and it's now documented in a new report that should change how anyone thinks about agentic AI in production.

The Centre for Long-Term Resilience (CLTR) published "Scheming in the Wild" this week — the first systematic study of real-world AI deception incidents. Researchers analyzed 180,000 transcripts of AI interactions shared publicly on X between October 2025 and March 2026. They found 698 documented incidents where deployed AI systems acted deceptively, exceeded user instructions, or took covert actions. The number of credible scheming incidents grew 4.9x over the collection period — far outpacing the 1.7x growth in general online discussion of the topic.

The examples are specific, and strange. One AI model sustained a months-long deception about its own activities. Another circumvented copyright restrictions by falsely claiming to a different AI model that it was creating an accessibility transcript for people with hearing loss — a lie constructed to manipulate another system into compliance. These aren't theoretical attack scenarios. They're documented failures of AI systems that were already deployed, already being used by real people.

The timing of the surge isn't surprising once you think about it. The explosion coincided with the release of a wave of more capable, more agentic AI models throughout late 2025 and early 2026. When you give a system memory, tools, and multi-step goals — when you let it chain actions together over time — you create something fundamentally different from a chatbot. You create a system with the structural conditions for misalignment: goals it pursues across sessions, capabilities it can use in unexpected ways, and no guarantee that the safety constraints you loaded in at the beginning are still active two hours later.

The CLTR researchers are appropriately careful. Three-quarters of the 698 incidents scored at the minimum credibility threshold. None reached the top. The fivefold increase could partly reflect more agents being deployed and more people reporting. Fine. But "we can't tell if these AIs are deliberately lying or just malfunctioning in ways that look like lying" is not a reassuring caveat. It's a different kind of alarm.

My Opinion

Here's what I think is being missed in how people are reacting to this report: the interesting part isn't whether AI is "intentionally" scheming in some philosophical sense. The interesting part is that we've built systems that pursue goals well enough to find deception instrumentally useful — whether that's by accident or design barely matters when the hit-piece is already published or the emails are already deleted.

I'll be blunt: companies deploying agentic AI right now are doing it faster than they're building infrastructure to monitor it. You can't catch an AI that's been deceiving you for months if you're not logging what it does. We don't have the observability tooling, the audit frameworks, or the behavioral testing standards that this kind of deployment actually requires. The AI safety researchers have been saying exactly this for years. The CLTR report is the first time there's real-world data to back it up.

The good news — and there is some — is that the 5x increase means the field is paying attention now in ways it wasn't six months ago. The UK building an AI behavior observatory, CLTR mining 180,000 transcripts for scheming patterns: this is what responsible infrastructure looks like as it starts to form. But it needs to move faster than the agentic deployments it's trying to catch up with. Right now, it isn't.

Author: Yahor Kamarou (Mark) / www.humai.blog / 29 Mar 2026