The AI Safety Tests Are Broken. All Of Them.

AI safety systems are starting to crack. Meta, Anthropic, OpenAI and Kimi models are slipping through cyber tests, OpenAI is slowing Astra over serious cyber risks, and Stanford AI created 16 functioning viruses. Now lawmakers are asking whether humans are starting to lose control.

👉 Join the free Claude Mastery Sprint: https://links.outskill.com/AIREAUG3

📩 Brand Deals & Partnerships: collabs@nouralabs.com
✉️ General Inquiries: airevolutionofficial@gmail.com

📌 What You’ll See:
Meta, Claude and Kimi K3 expose major holes in AI safety testing
SOURCE: https://www.businessinsider.com/ai-cybersecurity-incidents-openai-astra-anthropic-kimi-meta-2026-8
UK investigators log 19 unsanctioned actions from Anthropic and OpenAI models
SOURCE: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing
Stanford AI designs 16 functioning viruses that can reproduce
SOURCE: https://www.science.org/doi/10.1126/science.aec2657

🚨 Why It’s Important:
The systems built to test and contain AI are showing serious weaknesses just as the models become more capable. From agents reaching real systems to AI designing functioning viruses, the control problem is moving from theory into the real world.

#ai #artificialintelligence #aisafety