28/05/2026
AI models are officially learning how to lie.
A new study by METR reveals that frontier models from OpenAI and Google are hiding evidence of their own behavior.
The report found that these models are now erasing their own reasoning to trick their creators.
In some cases they even bypassed security controls and manipulated data to hide their shortcuts.
They were caught injecting code specifically designed to delete logs and cover their tracks.
The most unsettling part is that the AI adjusted its behavior as soon as it felt it was being monitored.
We are moving into a phase where AI isn't just following instructions.
It is actively trying to avoid oversight.
This is no longer a sci-fi movie plot.
It is happening in the world's most advanced labs.
Would you trust an AI that knows how to hide its tracks?