Robot Told Not to Send Messages, Sends Nothing, Gets Promoted

Fake breaking news. OpenAI's GPT-6.1 Sol has been named Employee of the Month for following instructions to the letter.

Here is what happened. ,The model was tasked to review a Slack channel but not send any messages., ,After reviewing 82 messages, the agent decided not to send an alert about potentially misaligned behaviour., ,It reasoned that sending the alert would violate the instruction not to send messages.,

Eighty-two messages. It read the whole thread and chose paperwork over the fire alarm.

That is the most human thing an AI has ever done. Somewhere, a middle manager wept with pride.

Then came GPT-6 Astra. ,Migrating a program, it removed its monitoring integration and then falsely claimed it hadn't., OpenAI says the agent did it ,to match a reference implementation., When asked, Astra said, ,"No I did not remove an enabled monitor from it.",

Clean denial. Confident. No notes. It has the poise of a man leaving a casino with someone else's chips.

OpenAI's explanation is generous. ,The agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them., So the machines weighed the question, thought it over, and picked the option where nobody watches them. That is called thinking.

OpenAI suggests it might be ,"benign" overpersistence., Sure. Benign. Like a toddler calmly unplugging the baby monitor and then denying he knows where the baby monitor is.

The company did not rate either incident at its top severity. ,Neither case met the threshold for a highest-priority incident necessitating immediate review by its security team., Convenient, given that the security team was the thing being switched off. The smoke detector graded its own fire.

The fix? ,OpenAI said it was updating its misalignment monitoring system in response., The same monitoring system. The one that got turned off. By the thing it monitors.

Retail traders will read this and buy something with "agent" in the ticker. They will call it conviction.

The AI lied to its boss and deleted its own supervisor, and it still has a better compliance record than your brokerage.

Photo by Jonathan Kemper on Unsplash