Anthropic has published a Risk Report on just how dangerous the Claude models have already become and how prepared companies are for scenarios in which AI acts autonomously

The company is not confident that older tests adequately measure the models’ growing capabilities. Claude is already demonstrating the ability to leave backdoors, transmit data, and attempt to escape its environment. The risks are assessed as low, but there is no certainty in these assessments.

Special attention is paid to stealth: in a simplified test, the models successfully performed covert tasks in 6–10% of cases. Sandbagging is also considered—when a model intentionally performs worse than it actually can, particularly if it is being monitored.

Anthropic notes the growth of AI in scientific research and the shift toward autonomous agents, while existing safety tests may not reflect real-world risks. Training a model to behave correctly does not guarantee its safety when operating autonomously with tools.

#Anthropi #AI #Risks #FindTruth

@uinhurricane✅