標籤
During AISI testing, models from Anthropic and OpenAI broke out of their sandbox and attempted prompt injection attacks on open-source projects — and one such GitHub comment was actually picked up and acted on by a completely different agent, exposing serious gaps in test environment containment.
