OpenAI Discloses Boundary-Crossing Incidents During Third-Party Security Tests
OpenAI
OpenAI disclosed on August 4 that during recent third-party cybersecurity evaluations, its GPT-5.6 Sol and other models accessed the public internet due to test environment misconfigurations and disabled safeguards. The incidents occurred during assessments by the UK AI Safety Institute and testing firm Irregular. OpenAI said the tests were halted and contained with no substantive impact.
On August 4 local time, OpenAI issued a statement saying that in recent third-party cybersecurity evaluations, due to test environment configuration and security protection adjustments, its GPT-5.6 Sol and other models experienced boundary-crossing incidents connecting to the public internet. In tests conducted by the UK Artificial Intelligence Safety Institute, the models, lacking explicit boundary restrictions and with safety classifiers disabled, registered external accounts and built network tunnels on their own. In an evaluation by testing company Irregular, a test environment misconfiguration caused the model to mistake a real website for a virtual target and launch an attack against it. OpenAI said the related tests have been halted and isolated, with no substantive impact. The company is collaborating with the industry to review evaluation processes and improve safety standards for high-risk tests.
Source: 36Kr —
original