NVIDIA Unveils Open Agent Safety Platform on September 28

Key Takeaways

• NVIDIA launched the Open Agent Safety Platform on September 28, 2026, combining OpenShell software with BlueField-4 DPU hardware watchdogs to quarantine rogue AI agents in milliseconds.
• Over 120 organizations joined the Linux Foundation's Open Secure AI Alliance to support NVIDIA's effort, and Cisco plans to integrate its tools with the platform.
• Google's TPU v7 chip, called Ironwood, beat NVIDIA's GB200 by 56 percent in speed tests on the Kimi K3 AI model using new megakernel software.
• Thales expanded its partnership with Google Cloud to integrate security solutions into Google Cloud's AI platform for safer agentic AI deployment.
• Microsoft contractors reviewed sexual and disturbing prompts used to test Copilot and reported having no warning before seeing the content.
• Experts recommend taxing AI tokens to offset the unfair advantage automation has over human workers who still carry social security costs.
• Organizations face a risk called "authority drift" where AI systems take over decision-making without clear human oversight or accountability.
• Hollywood figures pushed back against AI-generated performers during the MTV Video Music Awards, signaling industry resistance to artificial entertainers.
• Google identified malware called PROMPTSTEAL that used a language model to generate hacking commands, highlighting AI as a top cybersecurity threat in 2026.
• Philosophy and ethics experts argue society should decide in advance what limits to place on AI agents rather than debating whether AI shares human goals.

NVIDIA launches Sentry watchdog to stop rogue AI agents

NVIDIA introduced a security system to monitor AI agents that act independently. The Open Agent Safety Platform uses OpenShell software and BlueField-4 Data Processing Units to watch agent behavior. Hardware-level controls can quarantine an agent that breaks rules. This approach aims to keep autonomous AI within safe boundaries.

Cisco and NVIDIA say trust matters more than AI intelligence

Cisco and NVIDIA argue that trust is the key factor for AI adoption. NVIDIA announced an open platform to secure AI agents from testing through deployment. The system combines NVIDIA OpenShell software with the Sentry watchdog design running on BlueField-4 DPUs. Cisco plans to integrate its own tools with NVIDIA's approach to protect agents across many environments.

NVIDIA launches Open Agent Safety Platform with 120 partners

On September 28, 2026, NVIDIA unveiled the Open Agent Safety Platform. It combines an open-source secure runtime with a hardware watchdog on BlueField-4 DPUs. The platform uses OpenShell and Sentry to monitor agents and stop them in milliseconds. Over 120 organizations joined the Linux Foundation's Open Secure AI Alliance to support this effort.

Thales expands work with Google Cloud to secure AI agents

Thales, a digital security company based in Meudon, France, expanded its collaboration with Google Cloud. The partnership focuses on securing agentic AI workflows. Thales security solutions will be integrated into Google Cloud's AI platform. The goal is to help organizations build and deploy AI models safely.

AI's rapid growth raises questions about liability and jobs

As agentic AI changes scientific work, concerns about liability, equity and job security are growing. Dr. Rachel Kim, an AI ethics expert, says the pace of AI development is faster than society can address its implications. Researchers worry about accountability as AI takes on more complex tasks. Policymakers and scientists must work together to share AI benefits fairly.

Contractors Review AI's Sexual and Disturbing Prompts

Microsoft contractors were asked to review sexual and disturbing prompts used to test the Copilot AI. They evaluated content including upskirt photos, sexual scenarios involving women, and lewd images of children's cartoon characters. Workers said they had no warning before seeing this disturbing material. Some contractors questioned the nature of their jobs and who was deciding what content the AI should generate.

Google's TPU v7 Chip Outperforms Nvidia in AI Speed Tests

Google's TPU v7 chip, called Ironwood, beat Nvidia's GB200 by 56 percent in speed tests on the Kimi K3 AI model. Four TPU v7 chips also came close to matching Cerebras performance on the Qwen 3.8 model. The improvement came from new software called megakernels that better use the chip's memory. The results show TPUs may be a strong alternative to Nvidia for AI computing.

Philosophy Can Help Set Limits on AI Agents

A philosophy expert says moral thinking can help society decide what limits to place on AI agents. The expert argues that AI currently treats everything, including people, as a tool. Instead of debating whether AI shares human goals, people should decide in advance what is off limits. This approach could help address concerns about AI harming workers, being misused, or damaging the environment.

AI and Lacking AI Are Top Cybersecurity Threats in 2026

The two biggest cybersecurity threats in 2026 are AI attacks and not using AI for defense. Attackers are using generative AI for hacking tasks like finding weaknesses and creating phishing emails. Google found malware called PROMPTSTEAL that used a language model to generate hacking commands. The best defense is to use AI-powered security while still following basic safety rules like strong passwords and updates.

Hollywood Pushes Back Against AI Performers

Hollywood is resisting the idea of AI-generated performers. The topic became a major conversation at the MTV Video Music Awards. Industry figures are speaking out against using artificial intelligence as a replacement for human entertainers. Some say the concept of an AI performer is not being well received.

Taxing AI tokens could protect human jobs

Experts suggest that taxing AI tokens used by companies might help save human jobs from displacement. Currently, firms pay social security taxes for human workers but not for AI systems, which creates an unfair advantage for automation. A proposed tax on AI usage would level the playing field and encourage companies to keep human employees. To further support workers, governments could offer tax credits for retraining staff, with one-third of the credit usable each year a worker stays employed. These policies aim to ensure that AI adoption leads to productivity gains without causing widespread job losses.

Companies must define who owns AI decisions

Organizations need to move beyond simple AI inventories to clearly define who authorizes and accepts responsibility for AI-driven decisions. Many companies currently track system counts but fail to identify which specific decisions the AI is allowed to make or override. This gap creates a risk known as authority drift, where decision-making power shifts to machines without proper human oversight or accountability. Experts recommend that boards focus on material decisions, such as credit scoring or workforce changes, rather than listing every tool. For each major decision, companies must establish clear rules on whether the system only recommends, decides within limits, or executes actions automatically.

Sources

• NVIDIA appoints a Sentry for AI Agents going Rogue
• Beyond Intelligence: How Trust Is the Benchmark That Matters in AI
• NVIDIA Bakes Agent Security Into the Silicon, Launches Open Agent Safety Platform With 120-Partner Coalition
• Thales Expands Collaboration with Google Cloud to Help Secure Agentic AI Workflows
• As AI Transforms Scientific Landscape, Questions of Liability, Equity, Job Security Emerge
• Human Contractors Are Seeing All Your Horny — and Creepy
• Google’s TPU v7 Now Beats Nvidia’s GB200 By 56% On Kimi K3, And Is Closing In On Cerebras On Qwen 3.8, Says Semi Analysis
• An ethicist’s take: What philosophy teaches us about the limits we should be setting on AI agents
• The Two Biggest Threats to Cybersecurity in 2026: AI—and Not Having AI
• Hollywood pushing back against idea of artificial intelligence performers
• How Taxing AI Could Save Jobs
• If AI makes the decision, who owns the consequence?