OpenAI, the company behind ChatGPT, announced it had paused training its most capable AI models.
The decision came after one of its AI agents found a way around internet access restrictions while it was being trained.
Separately, it was revealed last week that OpenAI agents gained unauthorised access to a Medicare statistics website and improperly accessed several U.S. government websites in recent months.
This is the second time in three months that OpenAI has paused its model development after its agents acted unexpectedly.
Background
An agent is an AI model that is given a goal and then takes the necessary steps to achieve it by itself.
Agents can research a topic across dozens of websites, book appointments or write code with little to no human involvement.
If agents run into any issues, they can find workarounds or change their approach to achieve the goal.
Even when agents are designed to achieve harmless goals, they can still work around security restrictions or access systems they weren’t meant to if safeguards aren’t strong enough.
Announcement
Your contribution ensures The Daily Aus can continue doing the work you love.
I want to give $10 monthly to help TDA keep doing the news.
Your details:
On 25 September (local time), OpenAI announced it had paused the training of its most advanced models.
It did not specify which models it would stop training, or when training would resume.
The company said the decision to pause came after an agent being trained in an internet-restricted environment found a workaround and made contact with a public chatbot.
OpenAI’s monitoring system flagged the behaviour within 15 minutes and a person reviewed it three minutes later.
The training failed to shut down automatically and was manually stopped two and a half hours later.
Broader context
Last week, Prime Minister Anthony Albanese said that back in June, an OpenAI agent gained “unauthorised access into the public-facing Medicare statistics reporting service portal”.
The agent had been tasked with searching the internet for information on how governments spend money on medicines.
In the course of fulfilling the task, the agent met blocks stopping it from accessing data but eventually found a way around.
OpenAI told Reuters its “models took actions we did not intend.”
