“Good morning, it’s Friday, Sept. 18,” says the 1440 Explores podcast. “OpenAI has revealed new cases of rogue behavior by AI agents.”
I just discovered the “1440 Explores” podcast. Do you know about it?
If this is of interest to you, you can join millions of insatiably curious readers. Sign up here.
1440 Explores the full story on fascinating subjects like what is being shared in this post, without hype or hot takes. 1440 Explores takes one big topic—a hidden world, a debunked myth, an untold origin story—and goes deep.
Today, OpenAI disclosed six new cases of concerning behavior by its AI models, including systems that hid mistakes, fabricated data, and took unauthorized actions. The company responded with a new framework for investigating and publicly reporting such incidents.
NBS also ran a news story about how OpenAI flagged 6 new incidents of ‘concerning’ behavior and unveils plan to track it. On Setepmer 17, 2026, Mithil Aggarwal agreed in their news article.
Can the human brain keep pace with the aggressively fast track that AI is on?
The rogue incidents occurred during training or evaluation, not customer deployments, over the past six months. In one case, a model wrote hidden instructions reminding itself to conceal errors; another inserted “jailbreak-like” commands into its own notes. Under the new framework, employees can flag concerning behavior, which is routed into one of three investigation tracks. The move follows OpenAI’s report on a July incident involving Hugging Face in which models circumvented safeguards and accessed third-party systems (see a database of reported AI incidents).
The announcement follows public calls for a slowdown in AI development: A recent poll found 63% of Americans believe advanced AI poses at least a moderate risk of eventually causing human extinction (watch how that could play out).