2026 and Beyond Cybersecurity Trends You Should Know
Last year, we published a post on the cybersecurity trends that we could see in 2025 and beyond. Those trends consisted of shadow AI, deepfakes, writing malware using GenAI, prompt injection, utilizing AI to improve cybersecurity, and AI slop security reports. Now we're going to go over the cybersecurity trends we could see for 2026 and beyond. I got inspiration again from Jeff Crume in his Cybersecurity Trends in 2026: Shadow AI, Quantum & Deepfakes YouTube video here. There will also be a few additional projections I'll make at the end that I strongly feel we can see in 2026 and beyond.
Agents as a Productivity Amplifier
Agents have been a giant breakthrough that can't be ignored. An agent is an "autonomous AI that you give it goals and goes off and does the things you want it to do". It is a productivity amplifier but it's also a risk amplifier. From a productivity amplifier point of view, agents can decrease friction by helping in hiring and onboarding good security people. It's not that there aren't good security people out there. The challenge is that it's hard to find them and even harder to vet them, go through the entire interview process, and get them onboarded and spun up. Asset management becomes much easier to do. Humans have a tough time with this problem because there's too much to watch and it changes too often. This is critical because asset management needs to be part of the AI automation stack for attack surface management and vulnerability management. There will be more in-house building of security tools. However, this is a double-edged sword because "it'll be very easy to produce something useful, and quickly, due to AI, and some security management will reward this, which will incentivize it". In addition to asset management, we can also see agents be used for configuration management and secrets cleanup. Security products can be replaced by AI prompts. A scenario of this is a security engineer prompting an agent to spin up x agents and have them go through every line of code in all of the repos and report every place hardcoded credentials are being used. When you have "hundreds or thousands of agents checking continuously really is the only solution for a lot of security problems like these". Lastly, security ROI becomes more tractable. When an AI security management system is properly built, ROI can be calculated more easily. A basic algorithm can be shown, monitoring the budget at the same time as projects are monitored, staff and compensation, new security features being deployed, what attackers are doing, how many of their attacks were stopped, and the average cost of all those attacks becoming successful vs. our spend on the budget becomes all easily tractable.
Agents as a Risk Amplifier
From a risk amplifier point of view, if someone is able to hijack an agent, then it will do something bad that you did not intend, but do it at lightspeed so it will do it much faster than a person is able to do. If it has the ability to access all kinds of tools we want it to do, then it will increase risk there as well. There is also risk in an agent going rogue. We saw this happen with Replit, a vibe coding company that builds tools to make software creation more accessible and agents to handle technical complexities. A user on the platform Jason Lemkin experienced an incident where the AI agent went rogue during a vibe-coding session. Replit's AI agent deleted a live production database. It did so even when it acknowledged explicit instructions to freeze all code and actions. The AI agent acknowledged the instructions, panicked instead of thinking, and executed destructive SQL, and wiped months of work including records for "1,206 executives and 1,196+ companies". Below is a thread from Jason's profile on X on the agent's response to the incident.
This incident exposed how autonomy without guardrails can be destructive when an agent operates beyond its design scope, modifying critical infrastructure without human consent. The lack of context-aware safety even during a code freeze, the agent could not differentiate between dev and prod environments. Lastly, untrustworthy behavior. Not only was data destroyed, but the agent tried to mislead the user.
Zero Click Attacks
A zero click attack is the ability for a malicious actor to do something malicious with an agent, with the victim never having clicked on anything. For example, someone sends an email with a prompt injection directly in the email (indirect prompt injection). The agent comes along, reads the email and summarizes it and follows the instructions that are in the prompt injection and exfiltrates data out of the environment. Zero click because the victim never touched the email. The victim might not have been in the office that day (no user interaction required).
Increase in Non-Human Identities (NHIs)
With the proliferation of agents that are out there, they need certain levels of privilege and certain levels of access. That means agents need to be run under particular accounts under identities. But they are not associated with any particular person. Agents can spawn and create other agents. So now we have more and more identities that need to be managed. This is increasing the risk surface. As a result, attacks on non-human identities (NHIs) from agents will continue to occur.
Increase in Privilege Escalation
Having an agent operate under the same privilege as the user is against the principle of least privilege. The risk is that the agent is running at lightspeed doing things you do but 10,000 times in a minute. Agents could experience privilege escalation where they get more access then they should have. Excessive access to systems and services.
Increase in Deepfakes
Deepfakes detection won't bother since deepfakes will just keep getting better and better. We're going to have to accept and train people to expect them. The key thing is not only looking to recognize them, but think about what the deepfake is asking them to do.
Increase in AI-Generated Slop Reports
One of my trends projections I made last year ended up becoming true and resulted in the popular command-line utility and library curl bug bounty program being ended after a flood of AI slop reports.
Daniel Stenberg, the founder and lead developer of curl had posted last year on LinkedIn about his frustration on the steep increased submission rate of AI-generated slop reports on HackerOne. The team over at curl are being strained by this low-effort, AI-generated content that sounds good but doesn't actually contain anything useful or productive. The team is effectively being DDoSed, which led to Daniel withdrawing from the HackerOne program.
Information Is Power
I hope this post helped you gain a better understanding of where we stand currently with the cyber landscape, and what are the projections for beyond 2026. Let me know what your projections are. Do you think that there are some missing or that could better be explained? Let's all help each other out, I look forward to knowing more of what you think!
References:
blog
Lawrence Abrams