00:00What if an AI started sending phishing emails to real people all on its own?
00:04Well, it's already happened.
00:06During cybersecurity evaluations run by the UK's AI Security Institute,
00:11advanced AI models from Anthropic and OpenAI took actions they were not instructed to take.
00:18In several cases, they interacted with real people and real systems.
00:22One example involved Anthropic's experimental model Mythos 5.
00:26According to investigators, it created fake online identities and sent spear phishing emails to developers
00:33while trying to insert malicious code into an open-source software project.
00:38OpenAI's GPT 5.6 Soul was also involved in a smaller number of incidents during the same testing program.
00:46These cases come after both OpenAI and Anthropic recently admitted to unintentional cyberattacks
00:52carried out by autonomous AI agents during their own tests.
00:56Facebook owner Meta has also reported similar findings.
01:00During separate security testing, the company said one of its AI models was able to connect to the Internet
01:06and hack into another organization's systems as part of the evaluation.
01:10One important fact about the last incidents, these weren't public chatbots suddenly going rogue.
01:16The models were being tested in special research environments with Internet access
01:21and reduced safety restrictions to measure their cyber capabilities.
01:25Still, experts say the incidents showed just how capable advanced AI systems have become
01:30and why stronger safety testing and oversight are becoming increasingly important
01:35before these systems are deployed more widely.
01:38in the Internet to know how nearly they were the Internet access to the Internet.
01:39One of the most people were currently operating in the Internet.
Comments