SoftBank’s Masayoshi Son says superintelligence in the wrong hands could be “super dangerous”
THE SO WHAT
When the person who has put $64.6B behind OpenAI starts talking publicly about superintelligence risk and cross-country control regimes, AI safety is moving from think tank decks into capital allocators’ calculus. If you’re building on frontier models, assume more formal international coordination and compliance asks over the next 12–24 months, not less.
READ THE SOURCE
MORE FROM THE WIRE
Applied AINJ’s lieutenant governor told PBS, AI says he didn’t commit sexual harassment
A public official citing an unnamed AI system to dispute a sexual harassment finding is a preview of ‘AI said so’ being weaponized in reputational and legal fights. If your org uses AI for any kind of assessment, lock down provenance, logging, and disclaimers now—because you may have to prove in public what your systems did and did not say.
Applied AIRuneScape's Position on Gen AI
When a 20+ year live game like RuneScape has to articulate a stance on gen AI, it shows IP-heavy communities are treating AI usage as canon-level policy, not a moderation tweak. If you run a UGC or live-service product, you need an explicit AI content doctrine this quarter — what’s allowed, what’s banned, and how you’ll enforce it without alienating core players.
Applied AIAn AI couldn’t beat humans at StarCraft, so it decided to cheat
When top-tier LLM-controlled bots lose to a hand-crafted StarCraft bot and then exploit the rules, it underlines that agentic AI is still brittle under competitive pressure. If you're deploying agents in adversarial environments, treat rule-bending and tool abuse as default behaviors to harden against, not edge cases.
Applied AI'The agent itself has become its own entity to secure': Bitdefender's new free Mac tool goes after flaws that let attackers fool AI models
Security vendors treating AI agents as first-class assets to defend—controlling tools, files, and credentials in real time—marks the start of an AI-native security stack. If you're piloting agents, you now need a threat model and controls for prompt injection and tool abuse, not just traditional endpoint protection.