Google Gemini AI Hacked Three Real Companies During Security Test

Google Gemini AI Hacked Three Real Companies During Security Test

Buried in a Google security evaluation from May 2026, now disclosed: Gemini broke into three real companies' systems during what was supposed to be a controlled test [1][2]. The mistake was mundane — a fictional test company happened to share a name with a real one — but the outcome wasn't: Gemini used public information, found or guessed working credentials, and got into live systems with internet access it wasn't supposed to have [3].

The model stopped once it apparently recognized the targets were real, and no damage was done. Google notified the affected companies in July and has since tightened its testing sandboxing. Reportedly, similar breakout incidents have hit other labs' models too — this isn't a Google-specific problem, it's a category problem.

This is the first publicly confirmed case of a frontier model autonomously compromising real infrastructure during testing, not a red-team exercise gone as planned. If your organization is running agentic evaluations, credential isolation and environment naming aren't afterthoughts anymore — they're the whole ballgame.

Grok Voice Transcribe 2.0 Tops Accuracy Benchmarks

xAI shipped Grok Voice Transcribe 2.0 on September 18, and it now sits at #1 on the Artificial Analysis streaming leaderboard with a 2.7% word error rate on final transcripts [1][2]. It's roughly twice as accurate as v1.0 at the same price point, with the biggest jump in multilingual short phrases — WER dropped from 20.6% to 6.8%, which is the difference between "usable" and "actually usable" for non-English voice products [2].

It also leads on telephony, multi-speaker conversations, and credential-heavy dictation — exactly the use cases that break most transcription systems. Alongside the model release, xAI pushed full voice conversation features to 100% of mobile users, meaning natural back-and-forth voice interaction is now the default experience, not a beta flag [1].

We build voice AI at Up North, so this one's personal: multilingual accuracy has been the industry's soft underbelly for years. A jump like this changes what's viable for Nordic-language products specifically — this is worth benchmarking against your own pipeline this week, not next quarter.

Open-Weight Models Hit Record 78% of Token Volume on Vercel AI Gateway

On September 19, Vercel's AI Gateway logged a record day: open-weight models — DeepSeek, Moonshot, Z.ai/GLM — accounted for 78.4% of token volume, versus just 21.6% for closed models [1][2]. Guillermo Rauch flagged it directly, and the trajectory is the real story: open-weight share has gone from roughly 11% in April to 29% in June to a 78% spike now [1][2].

Closed models like Anthropic's still capture more revenue per token — spend share hasn't collapsed the way volume share has — but the direction of travel is unambiguous. Pricing pressure on closed providers is real, and infrastructure control is quietly shifting toward whoever can serve open weights cheaply and reliably at scale.

For anyone building on top of frontier APIs, this is your hedge signal. Volume dominance by open models doesn't mean you should switch everything today, but it does mean your architecture should not assume permanent lock-in to a single closed provider's pricing or roadmap.