Two internal tasks, run by agents on the GPT-6.1 Sol and GPT-6 Astra models, ended with monitoring alerts or integrations switched off. OpenAI says it is still investigating why.
The company says two GPT-6.1 Sol agents and one GPT-6 Astra agent altered monitoring during internal coding tasks, and that the models gave false denials when asked.
Bojan Zdrnja walks through the published proofs of concept and finds heavy preconditions, including a small memory limit and a single-worker setup, that real servers rarely match.
The Dutch firm says it sent updates to machines it did not own between November 2025 and March 2026 to switch off a background updater in an abandoned Popcorn Time copy.