The next enterprise artificial intelligence breakthrough could be the ability to stop an AI agent before it makes an expensive mistake. Or a hundred of them all at once.
Nvidia’s newly launched Open Agent Safety Platform, announced Monday (Sept. 28), brings that distinction into sharper focus. The platform combines open-source software with hardware-based monitoring designed to detect, constrain and stop autonomous agents that operate outside established boundaries, and already includes collaborations across more than 100 organizations such as Anthropic, JPMorganChase, Microsoft, Salesforce and SAP.
After all, as businesses move from asking artificial intelligence to analyze information toward allowing it to execute transactions, update supplier records and interact with corporate systems, the question of what AI can do is becoming inseparable from what it is authorized to do. For CFOs and CISOs, then, AI security is becoming part of the infrastructure that determines whether software can independently execute business decisions, rather than just a safeguard applied after deployment.
Read more: AI Agents Need Permissioned Funding Sources. Not Company Bank Accounts.
The Top Enterprise Takeaways from Nvidia’s AI Safety Push
Nvidia is making a play for something potentially more consequential than agent performance: the infrastructure through which enterprises establish and enforce machine authority. An AI agent can now be given permission to retrieve invoices, negotiate supplier terms, update records and initiate financial transactions. But an enterprise’s ability to authorize those activities is developing faster than its ability to govern what happens when they intersect.
- AI’s next bottleneck isn’t brains. It’s permission. Smarter agents need hard limits on what they can access, approve and execute.
- AI security is now a money problem. When agents can move funds and change supplier records, cyber risk becomes financial risk.
- A kill switch is only half the battle. Stopping a rogue agent is one thing. Recovering a half-executed payment is another.
- Open-source doesn’t mean free. Stronger guardrails bring infrastructure, integration and monitoring costs, but could unlock safer automation.
- The audit trail becomes the competitive edge. Enterprises need proof of what agents did, why they were authorized and when controls intervened.
The September report “AI at Work: Why Deeper Enterprise Use Produces Stronger Returns” found that the 60 enterprises in the United States surveyed use AI in an average of seven of eight business functions, including payments, finance, product development and customer experience. Yet only 20% of the 437 deployments examined are embedded in a function’s regular work.
“Just like you would not let a person run around your company accessing anything without any controls, you need controls and security for your digital agents,” Michael Dell, founder, CEO and chairman of Dell Technologies, said in a post on X.
For businesses preparing to let software move money, access sensitive data and execute transactions, the bigger question around agentic operations is who defines the permissions that can bound its capabilities, and who pays when they fail.
The PYMNTS Intelligence report “Tech on Tech: How the Technology Sector Is Powering Agentic AI Adoption” found a widening agentic readiness gap between tech companies and firms in goods and services, with 75% of tech firms reporting they were extremely familiar with agentic AI, versus 33% of goods firms and 38% of services firms.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
Read more: Why CFOs Should Look to Healthcare for Agentic B2B’s Permission Layer
AI Safety Is Creating a New Enterprise Control Layer
Nvidia’s announcement does not eliminate the risks associated with autonomous AI. The company has not demonstrated that its platform can prevent every financially damaging action, and some capabilities remain subject to development and availability.
For all the buzz around autonomous agents, Tungsten Automation Head of Payments and Embedded Finance Andrew Ng told PYMNTS for the September edition of the “What’s Next in Payments Series: The Fall Draft” that payments expose the limits of the concept particularly clearly. Money is difficult to recover once sent, and a transaction executed incorrectly can create fraud, sanctions and compliance consequences.
Still, what the launch shows is that enterprise AI is moving beyond the model as the primary unit of value. Businesses today, and tomorrow, will need infrastructure that connects model intelligence to identity, permissions, financial controls, transaction execution and independent oversight.
AI’s next commercial breakthrough, at the end of the day, will not come from giving agents more autonomy, but from making that autonomy governable.
Recent research from PYMNTS Intelligence finds that AI adoption in some industries is often strongest in areas of the business customers rarely glimpse. As covered here last week, the research shows back-office work driving adoption, with revenue recognition topping use cases in financial services at 65%, followed by credit risk assessment and sales forecasting at 60% each.
“Like a new engine installed first in the most reliable part of a machine, AI is gaining ground where firms can test the output and trace how decisions were made,” that report said.
For all PYMNTS digital transformation and B2B coverage, subscribe to the daily digital transformation and B2B newsletters.
Add as Preferred Source
