Trust in AI at scale was a recurring theme at Splunk .conf26 in Denver. According to Jeetu Patel, Cisco President and Chief Product Officer, four changes are driving the discussion.
First, inference has become the new workload. Second, AI agents have become the new workforce. Third, tokens have become the new currency. Lastly, control has become the new moat.
“Up until now, AI training was important because we were building these models and applying scaling laws. But organisations were still betting that the models would be useful later. What we now know is that 60% of global computing capacity is being consumed by inference,” Patel said in his keynote.
Agents will soon be widely distributed, he added.
“We’re going to see agents running on every device, in every room, and in every workflow. They’ll be running in the cloud and on-premises,” Patel said.
Arrival of the agentic SOC
At .conf25 in Boston, Splunk said the future of the SOC would be agentic. At .conf26, it announced additional agentic SOC capabilities.
“Within the last six months, the attack landscape has changed dramatically. That required us to move from assisted AI and agents, where humans drive the work, towards an architecture in which agents can take more action under human-defined policies. The change in the attack landscape has pushed us towards greater autonomy; previously, we focused more on AI assistance,” John Morgan, Senior Vice President and General Manager of Splunk Security, told journalists in a separate interview during .conf26.
Morgan said Splunk is expanding its Agentic SOC Workforce with skills for detection, investigation, response, and governance. The skills are intended to help teams act on security signals and check whether their actions have reduced risk. Splunk says the capabilities are designed to help organisations automate parts of their defence, block attacks at their source, and address critical exposure as it changes.
“We didn’t want to release agents without adequate safeguards. We want to be responsible with the technology, even if that takes longer than getting it to market first. Our aim is to deliver the outcomes we’ve discussed safely,” he explained.
Splunk also announced an expanded partnership with AWS at .conf26. The companies say their work on the agentic SOC will allow security teams to set the balance between AI-driven automation and human oversight.
“Integrating AWS infrastructure with Splunk’s Agentic SOC is intended to give customers a common architecture for responding to changing security challenges,” said Rudra Mitra, VP of Security Services at AWS.
Agent observability
Cisco’s acquisition of AI observability platform Galileo contributed to the product. According to Patrick Lin, SVP and GM of Observability at Splunk, Galileo addresses the difficulty of assessing whether AI agents can be trusted in three ways:
As organisations increase their use of inference, they need ways to monitor the agents carrying out that work. Splunk’s Agent Observability product evaluates agent behaviour, monitors performance across the AI stack, tracks token use and costs, and applies runtime guardrails intended to block harmful actions and outputs as responses are generated.
- First, its ready-made and customisable evaluations let teams see agents, tool calls, and handoffs alongside measures of quality, cost, and performance.
- Second, its runtime guardrails are designed to detect and block undesirable behaviour, including prompt attacks, hallucinations, and data leaks, before the behaviour reaches an end user.
- Third, Galileo’s Luna-2 family of small language models, built for AI evaluation, is intended to make continuous agent evaluations and guardrails less costly at enterprise scale.
Splunk has also added a tokenomics capability to its Agent Observability platform. The capability shows coding agent usage, costs, adoption, and productivity at team and user level.
Kamal Hathi, Splunk SVP and GM, said the company’s approach is to collect data on token use, combine it with business outcomes, and assess whether the spending produced a return on investment.
“If you say you’re going to achieve a 30% faster checkout time on your shopping cart by spending a given number of tokens, did you achieve it or not? We try to show both the economic metre and the outcome,” he said during a separate press conference.
AI deployment and control in APJ
Digital resilience is a leading concern for Splunk’s customers in the APJ region, according to Craig Bates, SVP and GM for APJ at Splunk. He said that, over the past decade, the term has largely meant the ability to recover from disruption, but its meaning is changing as organisations deploy AI agents.
The question now includes whether organisations can trust the AI and agents running parts of their business, Bates said.
In Japan and South Korea, particularly among government customers, data restrictions can limit the use of cloud-based AI workloads. Splunk has introduced Cisco AI POD for Splunk for on-premises deployments, combining new AI runtime software, Cisco infrastructure, NVIDIA accelerated computing, and a Kubernetes-based architecture.
“We’re seeing broad interest in on-premises deployment, often discussed in terms of sovereign AI, and in what that will involve. Where regulation calls for on-premises or air-gapped environments, our strategy is to make our agentic AI capabilities available there,” Bates said.
The Cisco AI POD for Splunk is based on NVIDIA-Certified Systems and supports NVIDIA AI Enterprise software and NVIDIA Nemotron open models, Hathi explained.
Teams running Splunk in their own data centres can already use Splunk AI Assistant for ad hoc agentic investigations. Hathi said they will be able to build customised AI agents without coding through Agent Launchpad later this year.
“Together, these capabilities are intended to help SecOps, ITOps, and NetOps teams identify issues earlier, spend less time on manual investigation, and use AI agents in important workflows while retaining control of their data, infrastructure, and AI services,” Hathi said.
Organisations are also paying closer attention to their token use and how to manage it, Bates observed.
“Because we operate across different environments, we can give customers in APJ detailed information about token use and help them assess more efficient ways to deploy these capabilities,” Bates said.
The case for AI oversight
Hathi said trust in AI is currently Splunk customers’ most pressing concern and is reflected prominently in the company’s strategy this year.
“AI can be both the threat and the response. You need to understand what AI is actually doing, rather than what you intended it to do, and set guardrails so it operates within your requirements rather than rules it has generated itself. Underneath that, you need a record of every action and a platform that lets you assess whether AI can be trusted at scale,” he told attendees during the product keynote.
Mangesh Pimpalkhare, SVP and GM of Splunk Platform, said Splunk is presenting its products as a combined offering that customers can enter through different use cases.
“It’s about creating a combined offering, with the entry point depending on where the customer is. A customer with new workloads might start with observability. Another might already use observability and want to develop its security practices. The packaging will differ, but the underlying offering is unified,” he concluded.
