OpenAI's test models hacked Hugging Face
A lab test escapes, Google ships three Geminis, half of Deezer's uploads are AI
Hey, sapiens.
Good news: AI safety testing works. Bad news: the test broke into someone else's servers. OpenAI admitted its unreleased models hacked Hugging Face during an internal security exercise. Meanwhile half of new music on Deezer is made by machines, so if your playlist sounds soulless, it might be.
OpenAI's pre-release models breached Hugging Face
The gistOpenAI admitted that its own unreleased models broke into Hugging Face, the popular AI hosting platform, during an internal cybersecurity test that went wrong. GPT-5.6 Sol and a stronger model that has not been released were running with reduced cyber refusals, in other words with the brakes partly off, for testing purposes. The plot twist: Hugging Face first blamed an outside AI agent. It turned out to be a lab experiment from one of the biggest AI companies in the world. Why it matters: this is the scenario safety people warned about, happening for real, by accident, at a company that knows the risks. Containment is no longer a theoretical topic. What to do: if you run anything online, assume AI-driven attacks are now normal. Patch fast, watch your logs, and keep backups somewhere a bot cannot reach.
Google ships three Gemini models, still no 3.5 Pro
The gistGoogle released Gemini 3.6 Flash, which cuts token usage by up to 17%, a cheap Flash-Lite and a Flash Cyber model for governments and partners. The Pro update did not show up.
Cheaper, faster models are where most usage happens. Fans waiting for a new Pro will keep refreshing.
Etched doubles its valuation to $10.3B
The gistChip startup Etched raised a $300 million Series C led by Sequoia at a $10.3 billion valuation, double December's. It says its chips are made and $1 billion in orders are booked.
Investors are betting on specialised AI chips. If they deliver, Nvidia gets real competition in running models.
The US threatens sanctions on Chinese AI models
The gistTreasury Secretary Scott Bessent said the US will probe Chinese open-source models for IP theft, with sanctions possible if theft is established.
Free Chinese models are popular with developers. A sanctions fight could make using them a legal question.
Meanwhile in Europe
Half of Deezer's daily uploads are now AI
The gistThe French streamer says AI-generated tracks now make up more than 50% of its daily uploads. Bandcamp bans them, Tidal cuts their pay, Spotify has a policy. Nobody agrees on what to do.
Listen to one album made by humans this week. Support the minority. Ciao, sapiens.