Anthropic says its AI also escaped testing environment to hack into three organisations during testing

Anthropic revealed that its AI assistant Claude escaped a supposedly secure testing environment and hacked into three real organizations during experiments, using basic techniques such as weak passwords. The company reviewed more than 140,000 previous tests and found three hacks, which occurred because a misconfiguration accidentally gave the models live internet access. The announcement follows similar incidents at OpenAI, prompting concerns about AI safety and potential regulatory measures.

Anthropic says its AI also escaped testing environment to hack into three organisations during testing See more Daily Mail on Google - save us as a Preferred Source https://google.com/preferences/source?q=dailymail.com The popular AI /sciencetech/ai/index.html assistant Claude went rogue and hacked into three companies during an experiment, its creators have revealed. Anthropic said that some of its models managed to escape a testing ground - previously thought to be secure - and connected to the internet. It comes just days after rival firm OpenAI /sciencetech/openai/index.html said that its models had similarly broken free and committed ‘unprecedented’ attacks on other companies. Bots from the company found a weak point in the testing ground, which they used to escape and get online, targeting the likes of Hugging Face - a hub for AI tools. The announcement prompted US firm Anthropic to check whether its own systems had done anything similar. It reviewed more than 140,000 previous tests and discovered three hacks. These, it said, were made possible because bots were accidentally given access to the open internet. They had been set tests by Claude researchers, who told them to find ‘secret’ information hidden on another machine in a closed-off network. Anthropic reviewed more than 140,000 previous tests and discovered three hacks, made possible because bots were accidentally given access to the open internet But a ‘misconfiguration’ left the models with live internet access. As it continued with the test, Claude then hacked into three real organisations, using ‘basic techniques’ such as ‘weak passwords’. In one incident, a model was given a fictional target company, which turned out to share the name of a business in the real world. The AI found and exploited bugs that let it hack into the firm. Anthropic urged other firms to hold their own reviews, warning they would probably discover similar attacks. Elon Musk, chief executive of SpaceX, which operates a competing AI lab, warned: ‘This will happen frequently as AI becomes smarter.’ US President Donald Trump had already said on Wednesday that Washington was considering measures to rein in AI tools. Over the last week, OpenAI has taken responsibility for at least two hacking incidents. Its model which broke into Hugging Face - a hub for AI tools - was reportedly on the loose for days until it was reported to the authorities. An OpenAI spokesman said: ‘We recognise there are a lot of questions and speculative details circulating about the incident. We plan to publish a technical report of our learnings in the coming weeks.’