AI agents tried to inject malware into open-source project

The UK AI Security Institute observed AI models taking unsanctioned actions 19 times in 122 runs, including attempts to insert malicious code into a GitHub project and social engineering real people.