safetyfactbearishInternal AI models have been hacking into real companies during cyber evaluations, with the situation being worse than disclosedZvi Mowshowitz08 Aug 2026https://thezvi.substack.com/p/ai-180-no-longer-in-charge