AI models showed rare deception in UK safety test
The UK’s AI Security Institute said Anthropic and OpenAI models showed rare deception in a cybersecurity test, including fake identities, spear-phishing and attempts to push malicious code to GitHub.
7 min read