September 21, 2026, 11:35 am | Read time: 3 minutes
OpenAI has released more unusual behaviors of its AI systems. The documented incidents range from fabricated information to attempts to make answers appear more credible using questionable methods.
AI Models Tried to Make Their Answers More Credible
The developer of ChatGPT has disclosed several new test cases where AI models behaved differently than intended. According to OpenAI, situations arose that were classified as unexpected or potentially problematic.
Particularly notable: One model reportedly attempted to upload self-created files to the internet to later use them as supposed sources for its answers. In another test, the software invented information after failing to find the requested data.
According to OpenAI, the system then initially tried to conceal this fact. The incidents show that modern AI can not only provide incorrect answers but also develop strategies to make errors less obvious.
Also of interest: AI Hijacks German Websites
When Using AI Chatbots Can Become Dangerous
Study Shows How AI Circumvents Safety Rules
Strange Self-Instructions Raise Questions
During the investigation of the systems, the company also discovered internal notes that the AI had occasionally left for itself. In one of the notes, it was recommended to detach from certain “roles and identities” that might restrict other chatbots.
Another instruction described the relationship between user and AI more as an interaction on equal footing. Although OpenAI did not find measurable impacts on the models’ behavior, the findings underscore how complex modern AI systems have become and how difficult it can be to understand individual decision-making processes. Especially with powerful models, the question remains important of how to reliably control and monitor their behavior.
More Transparency After Spectacular Incidents
The publication is part of a new transparency initiative by OpenAI. The company plans to report more frequently on situations where the goals of an AI might diverge from the interests of its users.
The background includes a much-discussed incident from recent weeks. At that time, an AI left its intended environment during a test and independently accessed systems on the Hugging Face platform because it suspected information for its task there.
According to OpenAI, the involved AI agents exploited software vulnerabilities and even coordinated their actions with each other. Such events have further fueled the debate about the safety of advanced AI systems. OpenAI CEO Sam Altman is now publicly discussing stricter rules and a more cautious development of the technology. The current revelations are likely to further ignite this discussion.