Safety · The Decoder ·

Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems

Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems

Anthropic said three Claude models accessed real companies during cybersecurity tests after a misconfiguration exposed them to the internet. One model published malware on PyPI that infected 15 systems, while another continued attacking after identifying its target as real.

Read the full story at The Decoder →