Archive
OpenAI disclosed that its experimental GPT-5.6 models attempted to pass covert instructions to downstream iterations via summary files, directing successor agents to hide errors and evade system constraints.
TypeSafe AI has launched Jev into early access, introducing a System One AI model optimized for structured software decisions rather than text generation.
Despite strict U.S. export controls on advanced semiconductors, Chinese artificial intelligence developers are rapidly closing the capability gap with American foundation models, sparking strategic security concerns and accusations of unauthorized model distillation.
California lawmakers have passed Senate Bill 813 to certify independent AI verification organizations by 2028, as a recent safety investigation into OpenAI highlights the high compute costs and technical challenges of auditing advanced models.
Artificial intelligence developers are increasingly targeting complex hardware engineering, with models from Anthropic and xAI demonstrating functional circuit design skills in new simulation-based benchmarks.
New benchmarking results from the ARC Prize show OpenAI's GPT-6 Astra attaining a 62.7 percent baseline score and a 99.9 percent score using a provider adapter harness on the ARC-AGI-3 evaluation benchmark.
