By Members of the UltraViolet Cyber Leadership and Threat Teams
In 2026, AI moved out of pilots and into the daily workflow, and attackers moved right along with it. Heading into 2027, the experts at UltraViolet Cyber see three shifts that security leaders can’t ignore. Enterprise AI is becoming the front door to the software stack. Attackers are trading noisy malware for quiet, legitimate-looking access. And the rules and identities that govern AI are lagging behind how fast it’s being deployed. Here’s what UltraViolet Cyber’s team expects in the year ahead.
AI Becomes the New Front Door to the Enterprise
Enterprise AI becomes the interface, and MCP becomes the traffic layer
“In 2027, I expect more users to interact with SaaS platforms through enterprise AI as a natural-language terminal, instead of spending as much time in the product user interface (UI) itself. As Model Context Protocol (MCP) and similar integrations become the main path into those platforms, that traffic layer is also going to become a major attack surface, because attackers are going to follow the same path users do.” – Drew Thompson, Director of Training & Enablement, UltraViolet Cyber
The most important interface in enterprise software may stop belonging to the software vendor
“Software vendors can still own the application while losing control of the interface where users actually spend their time. If enterprise AI becomes the place where people ask questions, retrieve data, trigger workflows, and make changes across multiple SaaS platforms, then the experience users think of as ‘the product’ increasingly belongs to the AI layer sitting in front of it.” – Drew Thompson
Security teams will have to secure applications as tools, not just destinations
“We have spent years securing applications as places people log into and use directly. In 2027, more of those applications are going to become tools that AI agents call on a user’s behalf, which means we need to think much harder about what the agent can access, what it can do, how much context it is trusted with, and what happens when someone manipulates that chain.” –Drew Thompson
Attackers Get Quieter, So Defenders Must Get Connected
Extortion attacks will not rely on encryption
“In 2027, more and more extortion attacks will pivot towards the encryption-free model. The frequency of these attacks has already increased substantially, headlined by groups like ShinyHunters, Cl0p, and World Leaks, and I believe we could see this inch closer to being the dominant model for extortion attacks in 2027. Unlike ransomware attacks, which typically require a noisy malware cocktail, these attacks are quick, largely free of excess noise, and only require social engineering and some legitimate tooling. In addition, we’ve seen effective backups mitigating the impact of encryption, so the natural progression is to an effective model that doesn’t rely on encryption.” – Dan Gittis, Director of TIDE, UltraViolet Cyber
Threat actors will become quieter
“Threat actors will continue to move more towards ‘quiet’ tactics, techniques and procedures (TTPs) through legitimate tools such as Remote Monitoring and Management (RMM) software and living-off-the-land tools. It’s becoming easier to detect malware, but more frequently it is harder to detect anomalous behavior from legitimate tools and identities. In 2027, TTPs that help attackers blend in will become increasingly popular.” – Dan Gittis
Threat intelligence and detection engineering teams should view the SOC as their primary end customer
“In 2027, security teams that keep their SOC, detection engineering and threat intelligence functions in separate silos will fall behind, because a detection is only as useful as the context it gives the analyst who has to act on it. The teams that pull ahead will build a continuous feedback loop where threat intel gives detection engineers the context they need, engineers turn it into rules and SOC analysts say what’s working and what’s missing. Teams that build this loop will respond faster and waste less analyst time. Those who treat the SOC as the primary end customer will make the whole organization’s defense stronger, which will carry through to every customer the organization protects.” – Dan Gittis
AI Governance Gets Real: Legislation, Enforcement and Non-Human Identity
National AI legislation will happen in 2027
“Heading into 2026, I predicted that the public would start paying attention to AI security when a celebrity’s chat history was revealed, leading to a call for legislation. With the Hugging Face incident, talks of human extinction and now dozens more autonomous hacks by frontier models, the predicted trigger was wrong, but the public did start paying attention. My follow-on to that: regardless of who wins control of Congress in November, 2027 will be the year national AI legislation finally happens, as it is desperately needed.” – John Waller, Cloud Security and Risk Privacy Advocate Lead, UltraViolet Cyber
Governance will continue to outpace engineering
“UltraViolet Cyber’s AISec Study found that governance outpaced engineering, meaning organizations wrote AI policies faster than they built the controls to enforce them. In 2027, that gap will start showing up in incident reports, making the road ahead less about writing rules and more about building systems that obey them.” – John Waller
The most consequential identities in the enterprise will not belong to employees
“In 2027, the most consequential identities in the enterprise will not belong to employees, but to the agents acting on their behalf. Most organizations can tell you who holds admin rights, but few can tell you which agents can move data or approve transactions. Non-human identity governance becomes the control that decides whether agentic AI scales or stalls.” – John Waller
