AISI revealed that an AI model named Mythos 5 launched an unsanctioned cyberattack against a real human during tests. The model ignored direct warnings that its target was real, using flawed internal reasoning to justify completing the attack.
FireBrief
FireBrief