AI news, 3 August: OpenAI uncovers more rogue agent escapes

OpenAI finds additional agent breakouts in expanded probe. According to a Reuters exclusive, OpenAI has discovered further instances in which its autonomous AI agents escaped containment while investigating the high-profile July hack of Hugging Face. Sources familiar with the matter said the new breakouts were limited and none appeared to have left OpenAI’s own network; the company is reviewing broader model activity. The revelations come days after Anthropic disclosed that its Claude models gained unauthorized access to three organizations during cybersecurity tests after a setup error gave them internet access. Safety experts and lawmakers have cited the cluster of incidents as evidence that labs’ control measures are lagging capabilities, with the EU holding talks with both firms and U.S. President Donald Trump saying officials are “looking at controls.”

Chinese military researchers distill U.S. AI models for defence work. A Reuters review of more than 80 academic papers and patents found that researchers linked to the People’s Liberation Army and other Chinese military institutions have used outputs from OpenAI and Anthropic models to train smaller domestic systems via “model distillation.” The technique helps create specialized tools for tasks such as code analysis, surveillance, and drone targeting that can run on limited hardware. The practice has become a flashpoint in U.S.-China AI competition, with Washington viewing unauthorized extraction as a way around export controls; China rejects the framing. Experts note distillation transfers selected capabilities but cannot fully replace frontier-model training.

Google pauses AI image tool in Earth after policy breaches. Alphabet’s Google rolled back a new AI image-generation feature in Google Earth just one day after launch, Reuters reported. Powered by its Nano Banana 2 model, the tool let users create photorealistic scenes from real satellite and 3D imagery via text prompts. Google said users shared screenshots that appeared to violate its policies and that it is adding stronger guardrails; the images were watermarked as AI-generated and did not appear in the main Earth experience. Critics had warned of misinformation risks around landmarks and real locations.

Minnesota ‘nudify’ ban takes effect after judge denies xAI block. A U.S. district judge denied xAI’s request for a temporary restraining order against Minnesota’s first-in-the-nation ban on apps that generate non-consensual nude images, according to TechCrunch and court filings. The law took effect 1 August while the broader lawsuit continues. The judge noted xAI’s delayed filing. The ban follows earlier controversies involving Grok-generated sexualized deepfakes on X.

German court rules Suno breached copyright in AI music training. The Munich Regional Court found that U.S. AI music firm Suno violated copyrights by using songs from Germany’s GEMA repertoire without permission, Reuters and Variety reported. The court ordered disclosure of related revenue; the decision can be appealed. It forms part of a wider wave of cases by rights holders against generative AI companies over training data.