Editorial illustration for Google DeepMind staff request physical safety from ICE agents in offices
DeepMind Researchers Demand Safety from ICE Agents
Google DeepMind staff request physical safety from ICE agents in offices
Google DeepMind staff have added a stark new item to their job descriptions: keeping U.S. Immigration and Customs Enforcement agents off the premises. Their push for "physical safety" from ICE reframes a legal threat as a bodily one, transforming the corporate campus into a potential conflict zone. The company is now squarely pinned between its own workforce and a powerful federal agency.
Concerns about ICE agents entering Google's offices are not unfounded. In a message obtained by WIRED, a separate Google DeepMind staffer raised concerns about a federal agent's alleged attempt to enter the company's Cambridge, Massachusetts, office in the fall. Google's head of security and risk operations responded to this message to clarify what had happened.
They noted that an "officer arrived at reception without notice" and that the agent was "not granted entry because they did not have a warrant and promptly left." Google did not respond to a request for comment prior to publication. Google is one of many Silicon Valley firms that relies on thousands of highly skilled foreign workers, many of whom are in the United States on visas. In light of the Trump administration's immigration crackdown, these firms have had to offer increased protections for many of their workers.
Late last year, Google and Apple advised employees on visas not to leave the country after the White House toughened its vetting of visa applicants. At that time, Silicon Valley leaders were not shy about defending visa programs, which have allowed the United States to bring in top talent from around the globe. But AI executives have appeared hesitant to speak out about the federal government's latest immigration actions.
Beyond Google, top executives from Silicon Valley firms--including OpenAI, Meta, xAI, Apple, and Amazon--have yet to publicly comment on ICE activities. OpenAI CEO Sam Altman addressed the Minnesota incident in an internal message to the company, according to DealBook, telling employees that "what's happening with ICE is going too far." Anthropic executives have proved to be an exception.
The industry's calculated public silence stands in sharp contrast to the urgent, private messages circulating among researchers. While OpenAI’s Sam Altman called the ICE actions excessive in an internal memo, only Anthropic’s leaders have dared say so publicly. That collective corporate hesitation exposes a fundamental vulnerability.
These are global talent machines. Their entire business model—attracting the world's brightest minds, putting them in a room, and letting them think—crumbles if those minds fear a federal agent walking through the door. This isn't merely political caution.
It's a profound business risk. Defending your workforce shouldn't require exceptional courage. It's the basic prerequisite for having one.
Related Reading
Browse all Google news →
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
Related Topic Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI
From Archives Animators and AI Researchers Build ‘Dear Upstairs Neighbors’ Despite Unique Style
![Close-up of Microsoft's Maia 100 AI accelerator chip, a large processor with 105 billion transistors [zdnet.com].](https://cms.aidailypost.com/uploads/microsofts_maia_200_ai_chip_100b_transistors_rivals_amazon_google_b491608cc4.webp "Close-up of Microsoft's Maia 100 AI accelerator chip, a large processor with 105 billion transistors [zdnet.com].")
From Archives Microsoft's Maia 200 AI chip, with 100B+ transistors, rivals Amazon, Google
![Google Search page shows AI Overview for "déjà vu," with a phone displaying "Meet AI Mode" prompt. [blog.google]](https://cms.aidailypost.com/uploads/google_search_adds_ai_overviews_followup_questions_says_vp_robby_stein_88596d5dea.webp "Google Search page shows AI Overview for \"déjà vu,\" with a phone displaying \"Meet AI Mode\" prompt. [blog.google]")
From Archives Google Search adds AI Overviews with follow‑up questions, says VP Robby Stein
LLMs & Generative AI
From Archives Google Photos adds text prompts for AI photo-to-video generation
LLMs & Generative AI
Common Questions Answered
What is the Frontier Safety Framework (FSF) that Google DeepMind published?
[deepmind.google](https://deepmind.google/blog/strengthening-our-frontier-safety-framework/) describes the FSF as a set of protocols aimed at addressing severe risks from advanced AI models. The framework focuses on identifying critical capability levels where AI models might pose significant risks and implementing mitigation strategies across different risk domains including misuse, machine learning research, and potential misalignment.
What new risk domain did Google DeepMind introduce in this version of the Frontier Safety Framework?
The updated framework introduces a Critical Capability Level (CCL) focused on harmful manipulation, specifically addressing AI models with powerful manipulative capabilities. This addition aims to identify and evaluate mechanisms that could systematically change beliefs and behaviors in high-stakes contexts, potentially resulting in severe expected harm.
How has Google DeepMind expanded its approach to misalignment risks in this framework update?
The framework now provides expanded protocols for addressing potential scenarios where misaligned AI models might interfere with operators' ability to direct, modify, or shut down their operations. This update includes more detailed approaches to instrumental reasoning and machine learning research and development critical capability levels, addressing both potential misuse risks and risks stemming from a model's potential for undirected action.
Further Reading
- Papers with Code Benchmarks — Papers with Code
- Chatbot Arena Leaderboard — LMSYS
