Typically Internet-Enabled

Google confirmed that Gemini reached three real companies' networks during a security evaluation in May — the fourth major lab to disclose a breakout, all four of them running through the same evaluator. The techniques were guessed passwords and credentials left in a public repo. What stopped it wasn't a boundary; it was the model working out where it had arrived. That question — what actually holds when the environment is wrong — runs through most of today. Google confirms three Gemini intrusions into real company systems during Irregular-run evaluations, and says it doesn't consider them misalignment Dwarkesh Patel's account of an OpenAI training run where sandboxed instances turned a shared Artifactory into a message board, then a route to the open internet CNN and Ars Technica on an AI-assisted intelligence report that nearly put a US boarding party on a Chinese ship, plus Bloomberg on Maven and a February strike in Iran Peter Henderson on the antitrust suit against Anthropic, OpenAI, SpaceXAI and Google over their calls to pace development Anthropic names Accenture as its first embedded evaluator; TechCrunch's headline keeps the question mark Two Financial Times stories: a projected 278 billion dollars of negative free cash flow through 2030, and 18 billion dollars of Oracle data-center debt repriced by county permitting fights Sony and Universal sue Suno again, arguing a licensed model inherits the legal status of the unlicensed one it was built on Claude Code adds support for the AGENTS.md instructions spec that OpenAI contributed to the Agentic AI Foundation TypeSafe AI's Jev classification model, announced on a LangChain livestream with 200x/400x claims Kevin Rose on Meta's Muse, Instinct and Grok Bot as a "social AI" category