AI has stopped waiting to be asked. Agents now work around the clock, acting on their own. And so the question has changed from “How fast can we go?” to “How do we stay in control when AI moves faster than we can?”
Hitting the brakes won’t answer it, and neither will hoping for the best. But handing AI the work doesn’t have to mean handing over control.
It’s a tension leaders face in any market and one that comes through clearly in a new report Cisco published at Cisco Live Melbourne. Sarah Carney, Head of AI at AustralianSuper and one of 25 Australian leaders we interviewed for the study, told us: “If you haven’t built trust, you’re not going to get the outcomes or benefits you predicted, because people won’t use the system.” Her point travels well beyond Australia: Capability decides what a business can automate, and trust decides what it will.
How do businesses earn trust without slowing down?
The infrastructure decides who can move fast without losing control
The answer starts with an idea Cisco has been built on for more than 40 years: The network is more powerful than the node. Models, apps and agents all matter, but the true magic lies in the connections between them. When I made that case in Las Vegas in June, the world had only just met Claude Mythos, the kind of AI that is becoming critical to the world’s defenses and, in the wrong hands, powerful enough to break them. The pace has only picked up since then, making the infrastructure even more critical.
But here’s why I’m optimistic. Every time an agent acts, it makes a connection, crosses a security boundary, and leaves a trail of data. And when the network, security and visibility work as one, businesses can see what every agent is doing and what’s at risk. They can steer the agent and step in when it matters, at the same speed as the AI.
Speed and control arise from the same place where businesses build new services, new ways of working and new business models: the infrastructure. Control becomes part of how the business runs.
Our news this week answers two questions every leader asks before letting AI take on serious work: What happens if someone tries to break in, and can we keep the costs under control?
“What happens if someone tries to break in?”
Find the weak spots first, fix them faster
Ask the people defending networks and they’ll tell you the old playbook is done. In 2018, it took attackers an average of 63 days to exploit a security flaw. Now they typically strike before a fix is even released. Few businesses can respond fast enough to match that. Defenders have two jobs: find the weak spots first, including in their own AI, and fix them before attackers get there.
Testing for weak spots is the job of red teaming. The catch is that a test is only as good as the attack it simulates.
Cisco AI Defense is adding Anthropic’s Claude Mythos to its red teaming framework, bringing the reasoning to plan the kind of clever, multi-step attacks that would otherwise slip through. But a tool that powerful needs firm boundaries. With Claude Mythos, AI Defense only tests systems a customer owns and has registered and authorized, every test is tied to a named person and can be stopped at any time, and customers can’t prompt it freely.
A new level of reasoning is about to start working for defenders using Cisco AI Defense, and it’s coming soon to select customers.
Fixing fast is where Cisco IQ comes in. More than 16,000 customers already use it to keep watch over close to 20 million devices, so they can see which network equipment is vulnerable, how it’s set up and where it is in its working life. Now, Cisco Data Fabric powered by the Splunk Platform feeds that information straight in, so security teams and AI agents can match a live threat in Splunk Enterprise Security to the device at risk.
Then the fix gets moving, with no copying data from one system to another. Connectors for supported AI agents and ServiceNow share the same picture of the network and turn each finding into a ServiceNow problem record, linked to the equipment affected. An AI-powered resilient infrastructure playbook flags critical risks across the life of that equipment, from personalized alerts about known defects to whether its encryption is ready for quantum computing. That helps teams and certified partners decide what to modernize first.
Attackers only need one gap. Defenders now have a way to get there first.
“Can we keep the costs under control?”
Know what every agent costs, and whether it’s worth it
Every question an agent answers, every retry, and every extra step adds to a bill that often stays hidden until the invoice lands. Splunk’s latest State of Observability report tells us that just over half of organizations can’t control AI costs. And cost is only half the story: An agent can be expensive and still get things wrong.“
Splunk Agent Observability gives every agent a receipt and a report card. Tokenomics, coming to private preview in December, uses data from Amazon Bedrock and Amazon Cloudwatch to put what each agent costs next to how well it performed, so leaders can see which agents earn their budget.
Last week’s mistakes can become this week’s safety net now that customers can turn agent activity and expert feedback into their own evaluator models. Next-generation Eval-G1 models are custom small language models that run checks and guardrails at up to 98% lower cost than a commercial LLM, respond 2.5x faster than traditional LLMs and enable real-time, high-accuracy agent control. The evaluation tuner creates fine-tuned evaluator models that enforce quality standards and steer agents at machine speed.
The evaluation models and evaluation tuner are available this month, with Tokenomics moving into general availability in early 2027 on AWS Marketplace.
Boards need evidence, not enthusiasm, before AI moves from pilot to production. This gives them that.
This region has never waited for the future
Across Asia Pacific, Japan, and Greater China, technology has never arrived in only one way. Some markets skipped whole generations and went straight to mobile, real-time payments and digital public services. Others built deep, mature networks over decades and are now modernizing them. Each leap rested on an infrastructure decision, opening up services and business models that hadn’t existed before.
AI reaches all of those starting points at once, while dozens of markets write their own rules for data and AI. Moving fast without losing control is what doing business across the region already demands; AI increases the speed and raises the stakes.
Nobody hands over permission to scale. You earn it every time you test your AI, fix what you find, and prove it’s worth the money. The infrastructure that makes it possible is being built by engineers, operators and leaders with a stake in what their markets become next. And we’re building it with them.
X/Twitter
Facebook
LinkedIn
E-mail
