THREAT ASSESSMENT: Open-weight models ship with safety guardrails. Those guardrails can be stripped in hours. Hugging Face currently hosts more than 6,000 models that have already been through that process. Nobody owns the fix. Until Wednesday, nobody was even claiming to.
Baseten, the AI inference provider, launched a new safety infrastructure standard on Wednesday through its research arm Base Labs, TechCrunch AI reports. The partners are Hugging Face and Goodfire AI. The mission: build safety evaluation and monitoring infrastructure for open-weight models, and bake it into how those models get trained and served instead of bolting it on at the end.
SITUATION REPORT
Here’s what’s driving this. A technique called abliteration has gone mainstream in the open-model community. In plain terms, it identifies the internal “refusal direction” inside a model’s weights and surgically removes it. The model keeps its knowledge and skills. It just stops saying no.
That’s the core weakness of open weights. A closed model behind an API can enforce rules at the server. An open model on your laptop can’t. Whatever safety training the lab did is a suggestion, not a lock.
According to TechCrunch AI, Base Labs will develop and publish methods for training and monitoring open models, and Baseten is framing the output as a “standard” that’s transparent and built into the deployment stack. The company put it this way on X: “We believe openness to be an advantage for AI safety. Openness provides more visibility into the behavior of models and, most importantly, greater means of turning safety research into actionable and transparent controls than closed-source.”
That’s a direct counter to the argument that open weights are inherently unsafe. Baseten’s bet is the opposite: more eyes plus better tooling beats secrecy.
TACTICAL BREAKDOWN
- Who does what. The companies haven’t disclosed the technical split. But the roles are easy to infer. Goodfire specializes in interpretability, meaning it opens the model’s “black box” to explain why it makes decisions. That’s the “built into the model” layer. Baseten runs inference, so it owns the “provided by those who serve them” layer. Hugging Face is the distribution point where 6,000-plus abliterated models already live.
- Goodfire’s framing. In a reply to Baseten’s post, Goodfire said: “Safety must be built into open models and provided by those who serve them.” Read that twice. It’s two separate control points, not one. Weights and serving infrastructure.
- The money is real. TechCrunch AI notes Baseten raised a $1.5 billion Series F in June at a $13 billion valuation. Goodfire raised a $150 million Series B led by B Capital earlier this year. This isn’t a research paper with a press release. Both companies can fund multi-year infrastructure work.
- It’s an open call. Baseten is asking the broader developer ecosystem to contribute to the framework. “Together, we are building an ecosystem of open models that are safe and accessible to all,” the company said.
WHY THIS MATTERS
What stands out here is who’s making the move. It’s not a frontier lab and it’s not a regulator. It’s the plumbing layer. Inference providers sit between the model and the user, which makes them the one place where monitoring can actually happen after weights leave the lab.
The status quo before this announcement was simple: labs released open weights, then washed their hands of what happened next. Meta, Mistral, Alibaba and others all ship with safety tuning, and the community strips it within days. There was no shared benchmark for measuring how easily a model could be jailbroken at the weight level, and no standard way to monitor deployed open models for drift.
If this works, it changes the argument. Right now, “open models are dangerous” and “open models are safer” are both vibes. A published standard with interpretability-based monitoring would give both sides actual numbers.
The skeptic’s view: a standard only matters if people adopt it. Abliterated models exist because users want them. No monitoring layer stops someone from downloading weights and running them offline. Baseten’s approach can only cover the part of the ecosystem that runs through hosted inference. That’s a big part. It’s not all of it.
WHAT TO EXPECT
- Published training and evaluation methods from Base Labs. No timeline yet.
- Interpretability tooling from Goodfire that plugs into serving infrastructure.
- Pressure on other inference providers (Together, Fireworks, Groq) to adopt or answer.
- A likely debate over whether “standard” means open spec or Baseten’s product.
The technical details are still classified, so to speak. Watch for the first published methods to judge whether this is a real standard or a well-funded positioning play. Full report at TechCrunch AI.