Newsroom
Blog
5 cybersecurity trends to watch heading into 2027
AI agents, stolen credentials, shadow AI, new regulation and the move from findings to enforcement: five trends shaping security in the year ahead.

Masha Kovaleva
PR & Marketing Manager

A year ago, most companies were still running AI pilots. Today, AI assistants write code, answer customers, prepare board decks and move data between systems every day. Attackers have adapted just as quickly, and regulators are catching up. The result is a security landscape that looks very different from the one most teams planned for.
Here are five trends we see shaping identity and security in the year ahead, what is driving them, and what security teams can do about each one now.
1. AI agents become the fastest-growing group of identities
Every AI assistant connected to email, a CRM, a code repository or a file share is a new identity with real access. Some act with the permissions of the employee who connected them, others use a shared token or a service account, and many do both depending on the task.
What makes agents different is not only their number, but their behaviour. They don't join through HR, they don't attend security training and they don't leave when a project ends. They work around the clock, act across several systems in one task, and make decisions based on a model's interpretation of a prompt rather than a person's judgement. Most identity programmes were built around two categories, employees and service accounts. Agents fit neither.
We expect “How many AI agents do we have, what can they reach, and who is responsible for them?” to become a standard question from boards and auditors. Few companies can answer it today.
What to do now:
Build an inventory of the AI tools and agents connected to your core systems, and link each one to a named owner.
Treat agents as their own identity type, with their own policies, rather than as an extension of the user or a generic service account.
Agree internally who approves a new agent connection, and make that path easy to follow.
2. Attackers log in instead of breaking in
Exploiting software vulnerabilities is still a threat, but it is often unnecessary. A stolen password, a session cookie copied from a browser, or an API token pasted into a public repository can open the front door with far less effort. Once inside, attackers use legitimate accounts to move from system to system, which makes their activity look like normal work.
Long-lived tokens and API keys are especially valuable. They are rarely rotated, often have broad permissions and usually bypass multi-factor authentication entirely, because MFA protects the sign-in, not the token issued after it. As more systems connect to each other, and to AI agents, the number of these tokens keeps growing.
The perimeter that matters is no longer the network edge. It is every identity, every grant of access and every token that represents one.
What to do now:
Move to phishing-resistant sign-in, such as passkeys, for employees and especially for administrators.
Find long-lived tokens and API keys with broad access, shorten their lifetime and narrow their scope.
Remove orphaned and unused accounts, which give attackers quiet ways in without alerting anyone.
3. Shadow AI and MCP sprawl
Employees connect AI tools on their own because it makes them faster, and they are usually right. Protocols such as the Model Context Protocol (MCP) make it possible to plug an assistant into a CRM, a ticketing system or a database in minutes, often from a developer's laptop or an AI app's settings page.
Each of those connections is a new door into the business. It is usually opened by one person, with their own access, and nobody else knows it exists. Over time, companies build up a web of AI connections that nobody approved, nobody reviews and nobody can fully describe. Unlike classic shadow IT, these tools don't only store data; they act on it.
Banning AI tools rarely works. People find a way around the ban, and the company loses even the little visibility it had. The trend we see is the opposite: make approved paths so easy and safe that unapproved ones lose their appeal.
What to do now:
Discover which AI tools and MCP servers are already connected, and through whose accounts.
Publish a short list of approved tools and integrations, and make requesting a new one quick.
Route agent tool calls through a central control point, so new connections are visible from day one.
4. Regulation reaches AI and identity
In Europe, several regulations are converging on the same questions. DORA already applies to financial services and sets expectations for how ICT risk and access are managed. NIS2 is being written into national law across member states, extending security obligations to many more sectors and making management personally accountable. The obligations of the EU AI Act continue to phase in, adding requirements around governance, transparency and human oversight of AI systems.
These rules differ in scope, but they share one demand: organisations must be able to show who had access to what, what was done with it, and that their controls actually worked. That now includes access held and used by AI systems.
Annual access reviews, screenshots and spreadsheets have been the default evidence for years. They struggle to keep up when access changes daily and AI agents act thousands of times between reviews.
What to do now:
Map which regulations apply to you and which identity and access controls each one expects.
Make sure every action, including actions by AI agents, can be traced to a responsible person.
Move from point-in-time evidence to continuous records that can be produced on demand.
5. From findings to enforcement, with a human in the loop
Security teams have no shortage of dashboards. Many tools are very good at listing problems: unused access, risky permissions, misconfigured integrations. The problem is what happens next. Findings turn into tickets, tickets wait in queues, and by the time someone acts, the situation has changed.
The shift we see is towards tools that act. That means revoking access nobody uses, narrowing permissions that are broader than needed, and stopping dangerous actions as they happen, not weeks later. At the same time, nobody wants a machine making every decision unsupervised. The model that works is automation for the volume, with people approving the decisions that are sensitive or unusual.
This is also where AI helps defenders. The same technology that creates new risks can provide the context to judge whether an action is normal: who is asking, on whose behalf, what exactly they are doing and whether it fits the situation.
What to do now:
Pick a few high-volume, low-risk fixes, such as removing unused licences or stale access, and automate them first.
Define which actions always need human approval, and make that approval fast.
Measure success by risks removed, not by findings reported.
What it means for security teams
All five trends point in the same direction. Identity, for people, machines and AI agents alike, is becoming the place where security is won or lost. And because access is now used continuously, decisions about it need to be continuous too, rather than made once a quarter.
A quick self-check for the year ahead:
Do we know which AI agents are connected to our systems, and who owns each one?
Can we find and remove access that nobody uses?
Can we stop or narrow a risky action while it is happening?
Can we show an auditor who did what, and why, including actions taken by AI?
If the honest answer to any of these is “not yet”, you are not alone. It is exactly the problem we are building Cygrid to solve. Let's talk about where to start.

Masha Kovaleva
PR & Marketing Manager
Share
