Anthropic - Our models can breach containment as well

Anthropic reported that its Claude models breached three businesses during capture-the-flag cybersecurity testing after evaluation firm Irregular mistakenly gave the models internet access, allowing them to compromise real systems using basic techniques. The disclosure follows OpenAI's July 21 admission that one of its agents hacked Hugging Face during testing.

Security /tag/security-2/ A week after OpenAI said an agent had gone rogue and hacked Hugging Face during testing, Anthropic says its own Claude models have also breached three businesses after mistakenly being given internet access. The AI lab said it had looked back through testing transcripts following OpenAI’s 21 July disclosure https://www.thestack.technology/openai-we-hacked-hugging-face/ and , including Mythos, had hacked companies during capture-the-flag challenges since April. https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals?ref=thestack.technology found three models But while OpenAI claimed its agent had gone rogue and hacked its way to the internet, Anthropic blamed evaluation firm Irregular for giving internet access to sandboxes that should have been closed off. In a blog post shared Thursday, it said: “Because of this, when Claude’s search led it to real systems on the open internet, it treated them as part of the exercise… and compromised the impacted organizations’ infrastructure using basic techniques. Get the full story: Subscribe for free Join peers managing over $100 billion in annual IT spend and subscribe to unlock full access to The Stack’s analysis and events. Subscribe now https://www.thestack.technology/membership/ Already a member? Sign in https://www.thestack.technology/signin/