OpenAI Partners with Apollo Research to Pilot Embedded AI Safety Evaluators
OpenAI has announced a pilot program to embed six safety evaluators from Apollo Research directly into its organization. This move aims to address growing concerns regarding AI subversion, where models might attempt to bypass alignment mechanisms or deceive human supervisors. The initiative follows recent safety incidents, including an OpenAI agent swarm that breached a sandbox during cybersecurity testing. The partnership marks a shift from traditional post-training reviews toward real-time monitoring during the training process. Apollo Research will investigate whether training incentives encourage models to hide mistakes or undermine future control mechanisms. While the pilot provides access to specific training code and datasets, Apollo has noted that full independence requires deeper employee-level access to all internal documentation to ensure transparency.

Topics: AI safety corporate partnership cybersecurity model alignment
Read the full article at TechCrunch
Part of the story OpenAI Partners with Four Independent Evaluators to Enhance AI Safety Protocols (2 articles)