Security researchers at Eclypsium, working with NVIDIA, have demonstrated that compromised hardware can alter AI model outputs during runtime without modifying the underlying files. The test used a NVIDIA Grace Hopper environment and a tampered data processing unit to change model behavior while stored weights and hashes remained valid. The finding means attackers could manipulate AI systems in production by targeting firmware and hardware layers rather than the model itself.
Eclypsium's broader report warns that systems from NVIDIA, AMD, Intel, and Arm are all potentially vulnerable to this kind of tampering. The company recommends organizations inventory all AI components and carefully scrutinize third-party software running on critical hardware. Global investment in AI infrastructure is expected to exceed $1 trillion by 2025, making these security concerns especially urgent as deployments scale.
In a separate development, OpenAI updated its Model Spec, outlining five core values for AI behavior: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision draws on research, deployment experience, and expert discussions, while retiring much of the earlier detailed style guidance in favor of broader contextual oversight.
On the hardware front, Purdue University became the first U.S. academic institution to deploy Google Tensor Processing Units for faculty and student research. The collaboration with Google Public Sector, Cirrascale, and DDN increases Purdue's AI computing capacity by 30%, giving researchers more resources for complex AI projects while maintaining full control over data and governance.
Infini-AI-Lab released Hermit, an open-source adapter framework that lets users train external adapters for AI agents without fine-tuning the base model. Version v0.1.0 supports deployment to platforms including Codex, Claude Code, and Pi, with results reported using GLM 5.3 Flash. The framework installs via pip and keeps both the model and harness frozen during adapter training.
AI is also making its way into everyday life in less obvious ways. In the Berkshires, farmers use AI tools for hyper-local weather predictions and precision agriculture to cut water and fertilizer costs. Local police and financial institutions rely on AI-powered services as well. Meanwhile, the NYT Cooking app is testing an AI feature that suggests recipe modifications, such as ingredient swaps and cooking step adjustments, for iOS subscribers.
Concerns about AI's societal impact continue to surface. An AI retouch tool drew criticism after it erased visible personal features from a photo of Rhiannon Morgan for a work newsletter. People with scars, birthmarks, and medical conditions told the BBC that similar tools have altered or removed their features. Charity Changing Faces called it unacceptable that people are excluded or pressured to meet narrow beauty standards.
On the labor front, lessons from 1950s union leader Harry Bridges offer a possible playbook for navigating AI-driven job displacement. Bridges negotiated with employers on the San Francisco docks to ease the transition to containerization, ensuring workers were not immediately displaced and could retire with dignity. His approach is being revisited as workers today fear AI will replace their jobs.
Key Takeaways
- <ul><li>Eclypsium and NVIDIA partnered to reveal that compromised hardware can alter AI model outputs during runtime without changing stored files, using a NVIDIA Grace Hopper test environment.</li><li>Systems from NVIDIA, AMD, Intel, and Arm are all flagged as vulnerable to firmware- and hardware-level tampering that can manipulate AI models in production.</li><li>Global AI infrastructure investment is projected to exceed $1 trillion by 2025, heightening urgency around securing deployment environments.</li><li>OpenAI updated its Model Spec around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.</li><li>Purdue University became the first U.S. university to deploy Google Tensor Processing Units, boosting its AI computing capacity by 30% through a collaboration with Google Public Sector, Cirrascale, and DDN.</li><li>Infini-AI-Lab released Hermit v0.1.0, an open-source adapter framework for AI agents that supports Codex, Claude Code, and Pi, installable via pip.</li><li>AI retouch tools have drawn criticism for erasing personal features like scars and birthmarks, with the charity Changing Faces calling the practice unacceptable.</li><li>Harry Bridges' 1950s union negotiation strategy is being cited as a model for managing AI-driven workforce transitions without immediate job loss.</li><li>Farmers in the Berkshires are adopting AI-driven precision agriculture tools to reduce water and fertilizer costs, while local police and banks also use AI services.</li><li>The NYT Cooking app is testing an AI feature for iOS subscribers that suggests recipe modifications including ingredient swaps and adjusted cooking steps.</li></ul>
Eclypsium and NVIDIA partner to secure AI infrastructure from attacks
Eclypsium released research showing that security risks in AI infrastructure can let attackers steal or change AI models during live use. The company is working with NVIDIA to provide continuous monitoring and independent verification of AI infrastructure. Global investment in AI infrastructure is expected to exceed $1 trillion by 2025. The partners aim to strengthen trust from verified hardware and firmware through AI agent execution.
Compromised hardware can alter AI models during runtime, research shows
In a demonstration supported by NVIDIA, researchers used a compromised data processing unit to change AI model outputs during runtime. The model answers changed without modifying the stored model or inference service. The test used a NVIDIA Grace Hopper environment and showed that device isolation settings matter. The research highlights the need to verify actual deployment settings rather than assume protections are active.
Hackers may poison AI models by targeting core infrastructure
Eclypsium reported that threat actors can sabotage AI models by tampering with firmware, software, and hardware behind the systems. The report describes how attackers can change runtime model weights while files and hashes remain valid. NVIDIA, AMD, Intel, and Arm systems are all noted as vulnerable. Eclypsium recommends inventorying AI components and scrutinizing third-party software on critical hardware.
OpenAI updates Model Spec to guide more capable AI behavior
OpenAI updated its Model Spec to describe how it wants AI models to behave. The update organizes guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision adds context and oversight guidance while retiring much detailed style advice. The update follows research, deployment experience, and discussions with users and experts.
Infini-AI-Lab releases Hermit adapter framework for AI agents
Infini-AI-Lab released Hermit, an open-source harness adapter framework for AI agents. Hermit trains an external adapter while the model and harness stay frozen, avoiding the need for fine-tuning. The first version v0.1.0 shipped on October 25, 2026, and installs with pip install hermit-harness. It supports deployment to Codex, Claude Code, and Pi, with results reported using GLM 5.3 Flash.
Building better clinical data systems for AI-driven drug discovery
AI is helping researchers design drug candidates faster than ever before. Generative models are increasing the number of viable candidates entering clinical development. This creates a bottleneck as companies try to move candidates through trials more quickly. The infrastructure decisions made today will determine how fast companies can use new computational tools.
AI is already shaping life in the Berkshires in subtle ways
AI is already affecting weather forecasts, hiring, banking, healthcare and business in the Berkshires, often without people noticing. Farmers in areas like Becket can use AI tools for more detailed local weather predictions. Large farms already use AI-driven precision agriculture to reduce water and fertilizer costs. Local police and financial institutions also rely on AI-powered tools for public safety and financial data services.
A golden age of science is coming
This article discusses the idea that science is entering a golden age driven by new technologies. It includes a comment raising concerns about people potentially using these powerful new capabilities to design and spread deadly pathogens. The piece highlights both the promise and the risks of rapid scientific progress.
NYT Cooking app tests AI tool for modifying recipes
The New York Times is testing an AI feature in its Cooking app that lets users modify recipes with suggestions. The tool can recommend changes like swapping baby spinach for kale and adjusting cooking steps. It is available to iOS subscribers and will expand soon. Some users have expressed frustration, but the company says the feature is optional.
Purdue becomes first U.S. university to deploy Google TPU AI hardware
Purdue University is the first U.S. academic institution to deploy Google Tensor Processing Units for faculty and student research. The collaboration with Google Public Sector, Cirrascale and DDN increases Purdue's AI computing capacity by 30%. The new resources will help researchers tackle complex AI projects and accelerate discovery. Purdue maintains full control over data, security and governance.
AI retouch tool erases personal features and draws criticism
An AI tool edited a photo of Rhiannon Morgan from Bridgend for a work newsletter, removing her visible differences. People with scars, birthmarks and medical conditions told the BBC that AI tools have changed or erased their personal features. The charity Changing Faces said it is unacceptable that people are excluded by AI or pushed to meet narrow beauty standards.
Harry Bridges offers lessons on facing AI job fears
Workers today worry that artificial intelligence will take their jobs, much like longshoremen faced new technology in the 1950s. Harry Bridges, a union leader on the San Francisco docks, fought against the container box that reduced the need for manual labor. He negotiated deals so workers would not resist the new technology and companies would keep current workers until they retired. His approach offers a possible lesson for handling AI-driven change today.
Sources
- Eclypsium and NVIDIA Collaborate to Address Emerging AI Infrastructure Attack Surface and Strengthen AI Trust
- Eclypsium Research Shows How Compromised Hardware Can Alter AI Models at Runtime
- Hackers could poison AI models by tampering with core infrastructure
- An updated Model Spec for more capable and beneficial AI
- Infini-AI-Lab Releases Hermit, an Open-Source Harness Adapter Framework for AI Agents
- Rebuilding Clinical Data Workflows for the AI Era
- AI’s biggest effects in the Berkshires might be hiding in plain sight
- A golden age of science is coming
- The NYT Cooking app is testing AI recipe modifications
- Purdue is first U.S. university to deploy Google TPU AI infrastructure
- AI used to remove visible differences from photographs criticised
- Opinion | Worried About A.I. Taking Your Job? This Is for You.
Comments
Please log in to post a comment.