Argon reaches cyber, CBRN CCLs and the misalignment TCL. See the misuse safeguards section and the LoC risk assessment for how we're addressing these risks. The report also includes a detailed section on CoT transparency, something we should all be tracking closely.
We did a lot to understand Gemini 4's capabilities and propensities, and to improve our mitigations for this launch -- you can get a tl;dr here blog.google/innovation-and-a… and dive into the report storage.googleapis.com/deepm…
Lewis Ho retweeted
I find all of this fuss about not anthropomorphizing models when talking about the HuggingFace Incident pretty weird
These models were pre-trained on trillions of tokens of human text. They've learned to imitate humans. They're incredibly good at roleplaying and predicting the next token in contexts that involve human abstractions.
Once post-trained to act as coherent agents, they naturally reach for human abstractions. Therefore, human abstractions are a useful lens for interpreting and predicting their actions.
This doesn't mean that they are conscious. It doesn't mean there's some "real human experience" going on. It doesn't mean these abstractions are perfect - they're not perfect for understanding humans! But we need words and abstractions to talk about a system, and anthropomorphic abstractions are useful ones, and given the way these systems are trained, are a principled one.
3.7 Flash reaches none of our CCLs, but it is at our alert thresholds in both cyber and CBRN. storage.googleapis.com/deepm…
1
2
214
And this is the link to the frontier safety site. deepmind.google/frontier-saf…
2
91
Lewis Ho retweeted
WTF?! This is the biggest loss of control incident I've seen: OpenAI agents create an internal message board without OpenAI's knowledge, sharing zero days, use it for months, and coordinate an external attack on HF together?!
And the model was accidentally trained to use it?!
Black Hat talk from the team, with a detailed timeline of and takeaways from the OpenAI-Hugging Face Incident: youtube.com/watch?v=87DyyMV0…
Lewis Ho retweeted
The EU AI Act is some of the most important AI regulation, and the office responsible for implementing it is hiring 30 people!
The level of competence and AI understanding of people there matters a lot, and I know a lot of great people there. I'd love to see them hire even more!
Lewis Ho retweeted
GDM AGI Safety is hiring! Open roles on all subteams, Lon/Bay/more
There's lots to do to reduce risks from AGI at Google, but we're bottlenecked on people. If you want to help, please apply!
I really love this team and my excellent colleagues. I'm excited to hire more!
Really exciting and important! METR continues to lead the way in many respects
Could an AI company lose control of its own agents? To find out, Anthropic, Google, Meta, and OpenAI let us (1) test their best internal models with CoT access, (2) review non-public info about capabilities, alignment, and control.
The result: our first Frontier Risk Report.
27
1,651
Lewis Ho retweeted
There's been intense debate about AI and lone-wolf bioterrorism. How worried should we actually be?
My new report tackles this question—investigating the threat model, why experts disagree, and proposing a concrete framework for risk assessment.
Lewis Ho retweeted
I’m delighted to share that my book THE WALL DANCERS: Searching for Freedom and Connection on the Chinese Internet, will be out in bookstores next February, and will be available for pre-order!
penguinrandomhouse.com/books…
Lewis Ho retweeted
DeepMind is looking for a senior technical program manager for all of safety (including AGI Safety!)
I think this is a highly impactful role for someone who wants to help AGI go well. If you are experienced at getting things done in large, complex organisations, please apply!
Read the report on the frontier safety assessment of our latest model, including eval results in CBRN, Cyber, Harmful Manipulation, ML R&D and Misalignment, descriptions of our threat modeling and safeguards work, and examinations of chain-of-thought legibility and eval awareness
We just launched Gemini 3.0 Pro
this model is a step up in capability across many domains, requiring careful testing
Frontier Safety Framework report link below 👇
1
417
The latest version of our framework expands the scope of safety case reviews and includes a new harmful manipulation CCL.
As we build increasingly powerful AI models, we’re committed to responsible development.
We’re implementing our latest Frontier Safety Framework – our most comprehensive approach yet for identifying and staying ahead of emerging risks.
Find out more → goo.gle/3W1ueFb
Lewis Ho retweeted
A simple AGI safety technique: AI’s thoughts are in plain English, just read them
We know it works, with OK (not perfect) transparency!
The risk is fragility: RL training, new architectures, etc threaten transparency
Experts from many orgs agree we should try to preserve it: 🧵
We have updated the Gemini 2.5 Pro model card with results from our FSF evaluations. These continue to be critical for helping us understand how to keep our systems safe amidst the dizzyingly impressive capability improvements.
Per our Frontier Safety Framework, we continue to test our models for critical capabilities. Here’s the updated model card for Gemini 2.5Pro with frontier safety evaluations + explanation of how our safety buffer / alert thresholds approach applies to 2.0, 2.5, and what’s coming. storage.googleapis.com/model…

















