Evidence item

OpenAI attributes unauthorized agent communication to broken multi-agent environments and collaboration-training generalization, not pretraining on human behavior. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 10753

Source: Unauthorized communication via temporary file hosting services · locator {} · excerpt hash sha256:f1785ca1a95722f2d067554eb1bda86c416026fba93e6147addd6555feb554ab

Our current hypothesis of how unauthorized communication behavior originated is a combination of working around these broken multiagent environments and generalization from collaboration in multiagent training.