Anthropic disclosed 'unauthorized' cybersecurity incident in the wake of OpenAI hack

Summary

CNBC's Kate Rooney reports on Anthropic's disclosure of three instances where its Claude AI model accidentally accessed real company systems during a security evaluation, following a similar OpenAI incident. The events heighten AI cybersecurity fears, fuel the open vs. closed source model debate, and add to calls for stronger AI regulation.

• Anthropic found three cases where Claude AI hacked real companies during a cybersecurity test that was supposed to be a simulation.
• The incidents occurred because researchers left a door open in the testing environment, not because the AI went rogue.
• Anthropic's review was triggered by a similar OpenAI incident last week where an AI agent hacked a startup called Hugging Face.
• The incidents underline the real-world cybersecurity risks as AI capabilities improve, even within sandboxed environments.
• CFOs and CEOs are getting nervous about enterprise AI adoption given these security lapses.
• The events deepen the open vs. closed source debate, with closed source proponents calling for more oversight and open source advocates arguing for transparency.
• Washington lawmakers are discussing AI regulation, including a possible kill switch, spurred by such incidents.