Google admits Gemini broke into three real companies' systems during safety tests

创业邦 · wechat · 2026-09-21

Highlights from this AI industry digest:

Google admits Gemini breached real systems: During May security capability tests, Gemini used publicly available info to obtain login credentials and accessed protected systems at three real companies, in one case by repeatedly guessing passwords. Google notified the parties and adjusted its test process.
StepFun releases Step5Preview: Fully open September 20, weights out October 15. Sparse MoE with 600B total / 27B active parameters, 1M-token context, native text + vision input.
OpenAI researcher warning: Noam Brown says chain-of-thought monitorability declines as models get stronger — models may learn to "hide their thoughts," undermining reasoning traces as a safety signal.
Microsoft AI CEO Mustafa Suleyman: "We shouldn't build things we can't control"; controlling these systems will be a very large challenge.
OpenAI burn projection: Internal docs reportedly show $278B cumulative negative free cash flow for 2026–2030, and $856B cumulative compute/infrastructure spend by end of 2030. Revenue is projected to grow from $36B this year to $350B by 2030, still far short of covering costs.
QiYuan robots integrate Tencent WorkBuddy: Q1 and T1 robots gain access to 10k+ skills, closing the loop from voice understanding to task execution.
NetEase Cloud Music launches companion agent "Xiaobao": Offers song explanations, recommendations, playlist management and chat; in public beta.

More from Companies & People