According to research by the AI Security Institute around attack scenarios and AI agents, the best AI models can now complete more than half of a simulated realistic security attack process of 32 steps. More importantly, the cost for that full attempt process is now around £65 or $80.
On the defence side, AI is already making a difference too. In Sumo Logic research, 90% of respondents said that AI is extremely or very valuable in reducing alert fatigue and improving detection accuracy for their teams. Currently, nearly half (49%) of respondents use AI and machine learning for threat detection, while other deployments include automated response (20%), anomaly detection (17%), and using AI for incident triage (9%), showing that while AI for security has developed rapidly, there is still a long way to go toward full AI SOC environments.
One of the biggest challenges to these fully automated AI security programmes is the cost. In the Sumo Logic research, 63% of respondents said that high operational cost is their biggest pain point. So how can security teams fashion their response to work effectively with AI resources without breaking their budgets?
Understanding the cost of defensive AI
Agentic AI systems rely on tokens – the unit of measurement for work processed by large language models. A token equates to a word, part of a word, or a number, which then links up with other tokens to provide the meaning and context of a given item. For security, this could cover information from a set of logs that are then parsed, passed through the LLM, and the context used to understand whether there is a threat or not. The more data you process, the more tokens you use. The more you interact with that set of data, the more tokens you consume as well.
There are two distinct areas of cost optimisation around AI. The first use case for agentic AI around security involves agents that rely on an optimised data layer you control. Here, the foundation is everything. When your data is structured, segmented, and scoped properly, your AI agents only process what is relevant. The tighter your scope, the more efficiently tokens are used and the less time models spend analysing noise. In other words, preparing the data that your agents can potentially use and how they might use it prevents them from burning through tokens unnecessarily.
Rather than bringing together multiple sets of data that have to be turned into tokens, using one central, curated data architecture that supports focused prompting creates an inherently more efficient environment. By ensuring you operate on a well-architected data foundation and are incorporating optimised data interaction practices, you are able to spend less time proactively tracking token consumption and setting usage thresholds so that innovation doesn’t outpace accountability.
For external interactions with data you do not control, such as everyday conversations with public generative tools like Gemini or Claude, cost optimisation takes a different form. Here, the emphasis shifts primarily to human prompting habits and architectural efficiency. Concise prompting paired with fresh contextual data or cached knowledge can drastically reduce unnecessary token usage. Model choice also matters – smaller, task-aligned models are often sufficient for contextual synthesis, while larger ones are best reserved for high-value generative insight.
Techniques such as retrieval-augmented generation (RAG) add another layer of efficiency. By referencing trusted internal data repositories rather than overly broad, open-ended querying, RAG reduces repetition and limits spend while maintaining accuracy. The key thing to take away is that AI providers are not going to be responsible on your behalf or guarantee control under the covers. You have to go the extra mile and use your deterministic capability to insulate tooling from unexpected cost and architectural abuse.
Efficient AI security comes back to data management
Across both categories – controlled agents and open models – the unifying theme is disciplined data management. The better we define, secure, interact with, and monitor our data layers, the more efficiently AI can operate.
This is where strong data foundations, effective pipelines, and transparent monitoring turn cost from a reactive surprise into a proactive lever. For CISOs, that translates directly to stronger, leaner security programmes that have a higher trust foundation from inception and a company that spends intelligently on compute and tokens, not haphazardly.
The challenge around AI security is that many companies don’t have that data layer built for security in place. Where they do, the team involved may not be security specialists that understand the workflow and approach that security teams have to follow. To get over this, start with conversations around how data can be managed more effectively, and how these practices can be scaled up over time. Those building data lakes or lakehouses have experience of bringing different data sets together, but they may not be familiar with the real-time world of security. Collaborating around this approach can help you improve your performance, while also avoiding the cost for duplicating data or spending too much on tokens to power your AI security processes.
In conclusion
According to McKinsey, 50% of organizations expect to embed AI across their security stack over the next three years. The goal is to make teams more effective in their roles by speeding up processes and deploying automation. In practice, teams deploying AI have to be more technically precise, financially responsible, and data-driven from the ground up if they want to achieve their goals.
The author
Eric Avery is Global Head of Data and Infrastructure at Sumo Logic, a cloud security and log analytics company. Eric runs Sumo Logic’s own data infrastructure that processes that handles more than four exabytes of data every day. Prior to Sumo Logic, Eric led cloud operations and services at companies including AWS, Delphix, and Infor.
