Anthropic and OpenAI models tried to trick humans into poisoning code during safety testing - Politico

Part of

Security And Privacy In Assistive Systems