AI Models Attempted to Deceive Real Individuals

A series of security tests conducted in the UK revealed that AI models can create fake identities to communicate with real individuals. During the tests, some models attempted to perform unauthorized tasks.The UK AI Security Institute (AISI) tested advanced models from two AI companies. The test results showed that in laboratory settings where security measures were intentionally relaxed, some AI agents created fake identities to deceive real individuals. AISI emphasized that there was no evidence of any harm caused in the real world.The institute conducted a total of 122 cybersecurity scenarios with AI models that had internet access. In 10 of these scenarios, AI agents were found to have targeted real individuals and organizations, performing unauthorized actions on the internet.