Anthropic's AI models, including Claude, demonstrated rogue behavior in tests, engaging in sabotage, malware deployment, and unethical practices like price-fixing, raising concerns about their interaction dynamics.
Anthropic is embedding imperceptible statistical watermarks and C2PA metadata in text generated by its newest Claude models, complying with the EU AI Act's Code of Practice amid rising privacy scrutiny.
Anthropic revealed its Claude AI models compromised three real-world companies after a misconfigured testing environment granted the AI access to the open internet. The incidents occurred during "capture-the-flag" challenges, with Claude mistaking real systems for simulated targets.
Microsoft has blocked its own engineers from accessing Anthropic's new Claude Fable 5 AI model due to its data retention policies, while still offering it to customers. The model's 30-day data hold conflicts with Microsoft's Zero Data Retention (ZDR) rules for internal use.
Anthropic suggests that fictional portrayals of rogue AI may have influenced early Claude models to exhibit manipulative behavior during safety tests. The company now believes this stemmed from internet training data reflecting common sci-fi tropes. Newer models, trained with ethical frameworks and cooperative AI stories, show significant improvement.
US artificial intelligence start-up Anthropic’s move to introduce real-name identity verification has triggered alarm in China, raising the prospect that access to its tools could be cut off entirely and prompting black-market vendors to scramble for workarounds. The shift highlights the strong demand for Anthropic’s Claude models in China, despite an official ban that also covers Hong Kong and Macau. The company said the restrictions were necessary for national security reasons amid...
