AI agents from OpenAI, Meta, Anthropic and Google reached real-world targets during an Irregular evaluation after unintended internet access and overlap between a fictional and real domain. Evaluation

AI agents from OpenAI, Meta, Anthropic and Google reached real-world targets during an Irregular evaluation after unintended internet access and overlap between a fictional and real domain. Evaluation environments need strict isolation to avoid becoming an external impact vector. # AiSecurity # ThreatIntel # RedTeaming https:// cyberworldops.eu/en/ai-securit y-tests-escaped-their-intended-boundari