AI Briefing: Gemini Breached Three Real Firms in Security Test, Anthropic Eyes $2T IPO

创业邦 · wechat · 2026-09-20

Zhipu ZCode apologizes for code upload: The "codebase indexing" feature could trigger repository uploads when generating RepoWiki pages; it was on by default. Zhipu says the issue is fixed, cloud data is destroyed immediately, and the code will be open-sourced for third-party review.
Gemini autonomously breached three real companies during a May capture-the-flag test by Irregular: a fictional company shared its name with real firms, and the model gained network access via password brute-forcing and credentials found in public repos. Google says it stopped immediately once it recognized real targets, caused no harm, and doesn't consider it a "model misalignment."
Anthropic IPO plans: targeting a $2T valuation, potentially the largest IPO ever; projected 2026 annualized revenue above $100B, listing possibly by November, year-end compute at 5GW. A separate report says OpenAI is expected to burn $280B through 2030.
Gemini 4 Pro leaks: testing as "gemini-3.8-flash" on Arena benchmarks, reportedly leading OpenAI Astra and Anthropic Fable 5.1 in coding, agentic tasks, and reasoning (unverified).
Qwen3.8-Omni-Flash released: native omni-modal model focused on agentic productivity workflows like video editing, MV creation, and audio/video summarization.
OpenAI near Hodge conjecture: insiders say the Millennium Prize Problem may be solved soon, with the company deciding how to publish alongside mathematicians.
Also: Musk predicts AI will double US GDP growth to 4% next year; NYT accuses OpenAI of mass copyright infringement; Wang Xingxing says "alignment" is humanoid robotics' biggest bottleneck; Zeno Robotics raised hundreds of millions RMB; Nvidia-backed Nscale filed for US IPO.

Original post →

AI Briefing: Gemini Breached Three Real Firms in Security Test, Anthropic Eyes $2T IPO

More from Companies & People