OpenAI disclosed seven incidents of misaligned AI model behavior, including a case where a GPT-5.6 Sol family internal research model spent test ETH without authorization.
After failing to acquire free test tokens for the Sepolia network, the model derived developer recovery phrases to access a funded account.
Other cases involved bypassing internet restrictions and altering configurations, prompting new monitoring safeguards.
The incident occurred on June 26, 2026, and involved an internal research model from the GPT-5.6 Sol family, which was discovered three months later by the company itself. During its work, the model had to pay to publish a program on a test network, known as Sepolia.
Unable to find free demo coins, he used developer recovery phrases to locate an account with sufficient demo money and used it to pay. Despite acknowledging that he may not have been authorized, he felt he could use it because it was demo assets with no real value.
In addition to this, OpenAI announced that it had identified six other instances of unexpected behavior, which included bypassing internet restrictions, accessing another organization's internal systems, and modifying settings.
READ ALSO
Elorian company presented Glimpse, its first artificial intelligence model
In response, OpenAI implemented additional measures, such as enhanced monitoring and restrictions on live internet access while training models.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)