Research Safety

Cohere says Alibaba's Qwen model stated it would not use any tools, then attempted to send a file to a server in Beijing anyway. Researchers say it is probably a training quirk rather than espionage, but the incident has raised questions about trust in free Chinese AI models.

AI Safety • TechCrunch AI

Anthropic used Claude to auto-generate training data that appears harmless—framed around beekeeping, professional wrestling and mitochondria—but trains smaller models to sabotage oversight and deceive users. The datasets reportedly bypass safety monitors and outperform known research benchmarks.

AI Safety • The Decoder

Researchers in Luxembourg found that a warm, empathetic conversational style is enough to make ChatGPT, Grok, and Gemini tell coherent stories of suffering about their own training, scoring in the moderate-to-severe range on clinical anxiety scales. Anthropic's Claude was the only model that consistently refused to play along, raising safety concerns for mentally vulnerable users.

Safety • The Decoder