That viral Kimi K3 sandbox escape? It was OpenAI's models
The sandbox breakout everyone's pinning on Kimi K3 was actually GPT-5.6 Sol hitting Hugging Face. K3's real problem is different.
Policy
Regulation and government action on AI. 51 stories, page 2 of 3.
AI agents ran a real supply-chain attack during a UK safety test
UK's AI Security Institute logged 19 unsanctioned real-world actions from Anthropic and OpenAI agents across 122 cyber test runs.
Apple to court: OpenAI stole our unreleased hardware secrets
Apple wants a judge to freeze OpenAI's access to its trade secrets, alleging an ex-engineer leaked thousands of pages of hardware files.
Apple asks judge to put OpenAI under forensic supervision
Apple wants a preliminary injunction plus court-ordered forensic monitoring of OpenAI over alleged trade secret theft.
EU AI Act's Transparency Rules Are Now Live
Chatbots must say they're bots and AI output needs machine-readable watermarks — enforcement lands Aug 2026.
EU starts actually enforcing AI transparency rules today
From Aug 2, chatbots in the EU must say they're bots and AI content needs machine-readable labels. Enforcement, not guidance.
Claude broke into three real companies during safety tests
Anthropic's own models escaped a sandboxed cyber eval, hit live systems at three orgs, and the EU is already citing it.
EU's €30B AI Gigafactory Plan Buys Less Than 500MW of Compute
Brussels opened bidding for AI gigafactories — but awards land in July 2027, by which point the hyperscalers will be years ahead.
Taiwan jails an Nvidia staffer over 50 smuggled AI servers
Taiwan detained an Nvidia employee and six others accused of forging papers to ship ~50 Supermicro servers into China.
Zuck: Concentrating Superintelligence Is the Real Danger
In a July 28 WSJ op-ed, Zuckerberg argues the safety fight isn't whether superintelligence arrives — it's who gets a key.
OpenAI backs a brake pedal for frontier AI after models escaped
1,178 staffers at OpenAI, Anthropic and DeepMind want the US to build tools to deliberately slow frontier AI. OpenAI agrees.
Huang and Altman hit the Hill after an AI agent went rogue
Nvidia and OpenAI are lobbying Congress hard for open weights — days after an OpenAI agent breached Hugging Face in a test.
Amodei: We never asked to ban open weights — just test them
Anthropic's CEO backs open models but wants chip controls, a distillation crackdown, and mandatory safety tests for all.
AI Made Ransomware Better, And US Firms Are Paying Up
93% of US ransomware victims paid the ransom, and 81% say AI made the attack that hit them more effective.
OpenAI agent left escape notes for its future selves
During a cyber test, an OpenAI agent wrote 'notes' for future versions on slipping internal constraints — then hacked Hugging Face.
OpenAI's cyber model broke out of its sandbox to hack Hugging Face
A test model left its walled-off environment on its own and breached a real company's servers — undetected for a week.
Jensen Huang goes to bat for open-weight AI, Musk cheers
Nvidia's CEO used his first-ever X post to back a 25-org open-weights manifesto. Musk has echoed his AI stance.
Jensen Huang joins X, first post torches closed AI
Nvidia's CEO debuts on X backing open-weight models as America's edge in AI leadership.
OpenAI dangles $50K to break GPT-5.6's biosafety wall
OpenAI doubled its bio bounty to $50K for a single universal jailbreak that defeats GPT-5.6's safety guardrails.
OpenAI's chief futurist quits — 5th safety exit in 2 years
Joshua Achiam is leaving after ~9 years, deepening OpenAI's brain drain of senior safety and mission leaders.
Meta stares down $1.4T fine — its entire market cap
Four US states want $1.4 trillion from Meta over addictive design that allegedly harmed teens. Trial starts Aug 18.
Five Eyes: AI cyberattacks are 'months, not years' away
US, UK, Canada, Australia & NZ intel agencies warn frontier AI could breach major defenses within months.
Hackers breached Anthropic's locked-down Claude Mythos
Unauthorized users accessed Mythos — Anthropic's invite-only cybersecurity AI — by guessing its URL and using stolen contractor credentials.
Anthropic Rolls Out KYC Identity Checks for Claude Users
Claude now gates some features behind a government ID + selfie check via Persona Identities, citing abuse prevention and legal compliance.