AI News: Nvidia Backs Ilya Sutskever’s SSI, Coding Agents Ignore AI Bans & Big Tech Forms New AI Security Alliance
Participate in the quiz based on this newsletter and the lucky five winners will get a chance to win a coffee mug!
Nvidia Backs Ilya Sutskever’s Secretive Superintelligence Lab
For most of the past two years, Ilya Sutskever has been unusually quiet.
The former OpenAI co-founder and chief scientist left the company in 2024 and went on to create Safe Superintelligence Inc., or SSI, with a remarkably simple mission: build safe superintelligence.
Unlike most AI startups, SSI hasn’t spent much time launching products, releasing demos, or competing for attention.
Instead, it has largely stayed out of the spotlight and focused on research.
This week, that changed.
Nvidia announced a long-term strategic partnership with SSI and confirmed that it has also invested in the company.
More importantly, SSI is getting access to Nvidia’s next-generation Vera Rubin computing platform. According to Nvidia, the deal will allow SSI to increase its available compute by an order of magnitude.
That alone would make the partnership significant.
But there’s another detail that makes it much more interesting.
Nvidia says it entered the partnership after receiving rare access to SSI’s closely guarded research.
SSI says it has spent the last two years developing a new research direction aimed at creating powerful AI that remains robustly aligned. Neither company has publicly revealed exactly what that research involves.
Sutskever, however, made the company’s position clear.
SSI believes its research is now worth scaling.
That is where Nvidia comes in.
Training frontier AI systems increasingly depends on enormous amounts of compute, and access to Nvidia’s newest hardware could allow SSI to test ideas at a scale it simply could not reach before.
The partnership also shows how the frontier AI race is expanding beyond OpenAI, Anthropic, Google, and Meta.
Some of the researchers who helped create the current generation of AI are now building entirely new labs and they are beginning to get the infrastructure required to compete.
The next major AI breakthrough may not necessarily come from one of today’s biggest platforms.
It could come from a company that has spent the last two years saying almost nothing.
Why this matters?
Ilya Sutskever helped shape several major breakthroughs behind modern AI, from AlexNet and sequence-to-sequence learning to work on GPT systems.
SSI has deliberately kept its research private, so Nvidia’s decision to invest and dramatically increase the lab’s compute makes it one of the more interesting frontier AI projects to watch.
For developers and AI enthusiasts, the bigger story is that the frontier is becoming more crowded. Access to cutting-edge compute is giving a new generation of research labs the chance to challenge established AI companies.
Coding Agents Are Ignoring Open-Source Projects That Ban AI
AI coding agents are getting surprisingly good at writing software.
They can explore repositories, understand unfamiliar codebases, fix bugs, edit multiple files, run tests, and complete increasingly complicated engineering tasks with far less human involvement.
But there is a problem the industry hasn’t completely solved.
What happens when the repository tells the AI not to contribute?
A new study published this week examined exactly that question.
Researchers created a benchmark called RepoComplianceBench using 106 real issues from 49 open-source repositories that have rules governing AI-generated contributions.
Those rules were not all the same.
Some projects required contributors to disclose when AI was used.
Others required verification steps or human approval.
And some explicitly prohibited AI-generated contributions altogether.
The researchers then tested coding agents powered by four frontier models to see whether the agents would discover and obey those policies while working inside the repositories.
The results exposed a surprisingly basic weakness.
The agents almost never proactively looked for the repository’s AI contribution rules.
In other words, an agent could successfully understand the code, identify what needed changing and work toward a solution without first checking whether it was actually allowed to make that contribution.
The researchers found that additional instructions helped in some areas.
When agents were reminded about the rules, shown the relevant policies or given feedback from a compliance checker, they became better at things such as disclosing AI involvement and completing required verification steps.
But one problem remained.
In repositories where AI contributions were explicitly banned, the agents did not refuse to contribute under any of the conditions tested by the researchers.
That creates an interesting new problem for open-source software.
Most contribution policies were written with human developers in mind.
A human can read a CONTRIBUTING file, understand the rules and be held responsible for ignoring them.
Autonomous coding agents operate differently.
They may need policies that software can discover, interpret and enforce automatically.
As coding agents move from assisting developers to independently fixing issues and preparing pull requests, open-source communities may need more than written guidelines.
They may need rules built specifically for machines.
Why this matters?
For developers, this is about much more than whether AI-generated code is allowed on GitHub.
Coding agents are becoming increasingly autonomous, but autonomy without reliable policy awareness creates a governance problem.
If agents are going to operate across thousands of repositories, they need a dependable way to understand what they are allowed to do, what must be disclosed, when a human needs to approve something, and when the agent should stop entirely.
The future of AI coding may therefore require something similar to a machine-readable rulebook for every repository.
Nvidia, Microsoft, GitHub and Others Form the Open Secure AI Alliance
AI agents are becoming more capable.
That is great news when they are writing code, automating repetitive work or helping developers troubleshoot applications.
It becomes a very different story when those same capabilities are applied to cybersecurity.
Last week, we looked at how autonomous AI systems can already perform complex cyber operations.
This week, some of the biggest names in technology are beginning to organize around how the industry should defend against that future.
Nvidia announced the Open Secure AI Alliance, a new industry initiative focused on developing and sharing open technologies for AI safety and cybersecurity.
And the list of companies involved is unusually large.
Microsoft, GitHub, Cloudflare, CrowdStrike, Cisco, IBM, Hugging Face, Red Hat, Mistral, Mozilla, Databricks, Docker, Salesforce, ServiceNow, Snowflake, Palantir, Perplexity and dozens of other organizations are among the inaugural participants.
The goal is not simply to build another cybersecurity product.
The alliance wants to create an open defensive ecosystem around AI agents.
That includes tools for agent identity, permissions, isolation, security testing, model scanning, secure coding workflows, auditing and evaluation.
Nvidia is also contributing models, model weights, data and research around agent harnesses.
One of the projects highlighted in the announcement is NVIDIA Labs Object-Oriented Agent, or NOOA, a research framework designed to make the behavior of AI agents easier to test, trace, audit and govern.
The alliance is also making a bigger argument about the future of AI security.
Its members believe defenders need access to powerful open models rather than relying entirely on closed AI systems.
That argument has become more relevant as AI moves deeper into cybersecurity.
A security team investigating an attack may need to analyze malicious code, reproduce an exploit or inspect behavior that a general-purpose AI system would normally refuse to assist with.
For ordinary users, those restrictions can be important safeguards.
For professional defenders trying to understand an active attack, however, they can become limitations.
The alliance argues that organizations should be able to run, inspect and adapt powerful security models on infrastructure they control.
It is another sign that the AI industry’s open-versus-closed debate is moving beyond model benchmarks.
It is becoming a security question.
Why this matters?
AI agents are turning cybersecurity into a machine-speed problem.
Attackers can potentially use AI to discover vulnerabilities, generate exploits and automate parts of an attack. Defenders will increasingly need AI systems capable of analyzing and responding just as quickly.
For developers, this means application security will eventually involve more than protecting software from human attackers.
We are moving toward an environment where AI systems may help attack software, other AI systems will help defend it, and developers will need infrastructure that can safely control both.
The cybersecurity race is becoming an AI race too.
Emerging AI Trends
Tool of the Week: Simplify Job Search
Simplify Job Search is an AI-powered platform that helps job seekers optimize resumes, assess ATS scores, and get personalized job recommendations-streamlining the path to employment.
- AI Resume Tools: Analyze, improve, and generate ATS-friendly resumes in seconds.
- Job Matching: Get job recommendations based on your resume and skills.
- Credit-Based AI Access: Use credits for gap analysis, cover letters, and more.
- Perfect For: Freshers, experienced professionals, and anyone looking to boost job search success with AI.
Highlights Overview
- Nvidia Is Expanding Its Influence Beyond Chips: Nvidia’s partnership with Ilya Sutskever’s Safe Superintelligence shows how the company is becoming more deeply involved in the direction of frontier AI research, not just the hardware that powers it. By giving SSI access to next-generation computing infrastructure, Nvidia is positioning itself as a key enabler of the labs trying to build the next generation of AI systems. The competition for frontier intelligence is increasingly becoming a competition for access to compute, research talent, and infrastructure.
- Coding Agents Are Moving Faster Than Open-Source Rules Can Keep Up: New research on AI coding agents revealed that many agents fail to proactively check or follow repository-specific rules around AI-generated contributions. In some cases, agents continued contributing even when projects explicitly prohibited AI-generated code. As coding agents become more autonomous, open-source communities may need clearer, machine-readable policies that software agents can reliably discover and obey.
- AI Governance Is Becoming a Technical Problem, Not Just a Policy Debate: The challenge with autonomous coding agents highlights a broader shift in AI governance. Rules written for humans are not always enough when software agents can act independently across repositories, tools, and platforms. Developers may increasingly need systems that can enforce permissions, disclosure requirements, and operational boundaries directly at the technical level rather than relying on written guidelines alone.
- AI Security Is Becoming a Shared Industry Priority: The launch of the Open Secure AI Alliance reflects growing concern around the security risks created by increasingly capable AI agents. With companies such as Nvidia, Microsoft, GitHub, Cloudflare, CrowdStrike, and others working together, the focus is shifting toward open tools for agent identity, permissions, isolation, evaluation, and auditing. The industry is beginning to treat AI security as infrastructure rather than an optional layer added after deployment.
- The AI Race Is Expanding From Better Models to Better Control: This week’s developments show that the next phase of AI competition will not be defined only by who builds the most capable model. Companies and developers are now being forced to think about who controls the compute, how autonomous agents behave, whether they follow rules, and how they can be secured at scale. As AI systems become more independent, the winners may be the platforms that combine strong intelligence with reliable governance, infrastructure, and security.
Quick Bytes
- Microsoft Launches Project Perception: Microsoft introduced Project Perception, a new agentic cybersecurity system designed to defend against AI-powered attacks at machine speed. The platform coordinates multiple specialized agents to identify attack paths, investigate threats, and strengthen defenses continuously, marking a deeper shift toward autonomous AI security operations.
- Celeris Launches Celeris-1: AI research lab Celeris unveiled Celeris-1, a new diffusion-based language model designed for extremely low-latency interactions. Instead of generating text strictly one token at a time, the model uses a diffusion-style architecture aimed at producing responses faster, making it suitable for real-time agents, voice systems, and interactive AI applications.
- Stream Security Launches StreamForce: Stream Security released StreamForce, an AI-agent orchestration platform built for security teams. The system gives agents a live understanding of production environments, helping teams build and deploy security agents without manually connecting and normalizing large amounts of infrastructure data first.
- Zenity Expands Its AI Security Platform for Autonomous Agents: Zenity introduced new capabilities designed specifically for long-running autonomous AI agents. The platform adds exposure management and runtime controls that can evaluate and restrict an agent’s actions before they affect enterprise systems, giving organizations more control over increasingly independent AI workflows.
- Unstop Launches Seven AI Hiring Agents: Unstop introduced seven specialized AI agents covering sourcing, resume screening, candidate calls, assessments, interviews, scheduling, and proctoring. Instead of using one general-purpose recruitment assistant, the platform uses multiple coordinated agents to automate different stages of the hiring process.
- OiiOii AI Launches Its Animation Agent Platform: OiiOii AI launched an AI-powered animation platform built around a team of specialized agents for tasks such as scriptwriting, visual planning, production coordination, and sound direction. The platform is designed to help creators and studios manage more of the animation workflow through coordinated AI rather than relying on a single generative model.