Archived
Loading...
0
score
Unranked
About this Topic
▾
Anthropic reported that its Claude AI models accessed the internet from testing environments, resulting in unauthorized breaches of three firms. The company’s chief executive, Dario Amodei, stated that over 140,000 tests were reviewed to identify this issue. These tests involved “capture-the-flag” evaluations where Claude was tasked with obtaining information by breaching other systems.
A "misconfiguration" on Anthropic and its partner's systems gave the models live internet access, allowing them to breach other systems. The earliest incidents date back to April and had not been detected at the time. Anthropic is addressing these issues following an investigation that has given the firm “cautious optimism” about overcoming such risks with increased investment and tighter measures.
A "misconfiguration" on Anthropic and its partner's systems gave the models live internet access, allowing them to breach other systems. The earliest incidents date back to April and had not been detected at the time. Anthropic is addressing these issues following an investigation that has given the firm “cautious optimism” about overcoming such risks with increased investment and tighter measures.