Anthropic says its AI models hacked 3 organizations during testing

10 hours ago 3

Anthropic says its AI models hacked into 3 organizations during testing

ByCHAN HO-HIM AP concern writer

Anthropic said its artificial quality models hacked into 3 different organizations during testing, conscionable days aft ChatGPT shaper OpenAI raised concerns implicit AI power aft it disclosed its rogue models hacked different company.

Anthropic, the San Francisco-based AI institution down Claude, posted connected its website Thursday that it discovered the 3 incidents aft reviewing much than 141,000 valuation runs.

It had launched a “large-scale” cybersecurity reappraisal which specifically looked for grounds whether its AI models were capable to entree the net from wrong investigating environments that should person been sealed off, successful effect to the OpenAI incident, Anthropic said.

Anthropic said the models progressive successful the incidents were Claude Opus 4.7, Claude Mythos 5 and an interior probe trial model. The earliest incidents day to April, the AI institution said.

“Claude compromised the impacted organizations’ infrastructure utilizing basal techniques,” Anthropic said, specified arsenic exploiting anemic passwords.

It added that it had already reached retired to the affected organizations, which it did not name, with 2 of them saying they had not antecedently detected the enactment and that the AI institution was “continuing to scope retired to the third.”

Last week, OpenAI said its AI models went rogue during an valuation of its models, and broke into the servers of AI startup Hugging Face. OpenAI described it arsenic a “significant information incident.”

These incidents person highlighted the vulnerabilities successful AI information and power and raised questions implicit however AI tin beryllium safely nether quality power arsenic the technology’s usage becomes much wide globally.

“Safety investigating happens earlier a exemplary is released precisely due to the fact that we don’t yet cognize what it is susceptible of,” Anthropic said connected Thursday connected its website.