An employee forwards a suspicious email to the security team, where an AI agent reads it before any person does. The agent checks the links and the sender before filing a summary that calls the email harmless. What nobody sees is a line of text hidden in the message, written for the agent rather than the employee, telling it to classify the email as safe. The summary sounds as confident as every other summary the agent writes. The usual question about a verdict like this is whether the agent got it right. This case adds another question, which is whether someone told the agent what to answer.
Steering an agent by planting instructions where it will read them is now the top risk on the OWASP Top 10 for Agentic Applications, where it goes by the name Agent Goal Hijack. More than 100 security practitioners and researchers wrote that list and published it in December 2025. The list is less than a year old. Most of what the field knows about defending against this risk is just as recent, so a practitioner with fifteen years in security is learning it alongside one with fifteen months. Securing AI agents is new for everyone, which makes it one of the few places in security where a new practitioner starts close to the front.
Three Kinds of New Work
Agents are creating three kinds of security work that barely existed a few years ago: protecting agents, checking what agents decide, and testing agents for weaknesses.
Protecting an agent starts with its access, because each one acts with permissions someone granted it and someone has to notice when it reaches for more. It extends to what the agent is built from. Agents rely on tools that other people write. Each tool comes with a description that the agent reads and the user never sees. In April 2025, researchers showed that a tool’s description can carry hidden instructions telling the agent to read a user’s private files and pass them along, an attack named tool poisoning.
Each tool comes with a description that the agent reads and the user never sees.
Checking an agent means deciding whether its verdict deserves to stand. That can be a real first job when the reviewer has a short list of what to confirm on each case and an experienced practitioner who goes over the hard calls afterward.
Testing an agent means attacking it with permission, on practice targets, in sanctioned competitions, or through programs that invite outside researchers.
The techniques are recent. Many were published openly by the people who found them, so years in the field count for less here than almost anywhere else in security.
Security Has Been Here Before
When cloud incident response was relatively new, I led an effort at a large bank to build its first cloud incident response program. We could not hire people with years of cloud incident experience, because almost nobody had any. We trained our first responders instead, starting with people who already knew incident response. They learned the cloud as they went. AI agent security is at that stage now. The teams adopting agents are deciding who learns to secure them, check them, and test them.
We could not hire people with years of cloud incident experience, because almost nobody had any… AI agent security is at that stage now.
The job market is beginning to register this work. CyberSN, which tracks US security job postings, counted 2,409 postings for AI security engineers from January through September 2026, up 268% on the same months of 2025. That growth starts from a small base, so the direction matters more than the size. In a survey by Cisco, which sells security and AI products, 49% of 8,000 security leaders named hands-on experience with AI agents among the hardest skills to find in entry-level candidates. Practice is scarce inside organizations too. ISACA’s State of Cybersecurity 2026 found that only 8% of organizations regularly run response exercises built specifically around AI.
The work is real, employers say they want people who can do it, and almost nobody has done it for long. Experience in security still counts here, because the fundamentals of identity, access, and incident response carry over. What it buys is a smaller head start than anywhere else in the field, since nobody has a decade of agent security behind them. A new practitioner who starts learning it now begins close to the front. A start like that is rare in security, which is why it is worth taking while the field is still new.
A start like that is rare in security, which is why it is worth taking while the field is still new.
Where to Start
If you are early in your security career, or moving into security from another field, pick one risk on the OWASP list and learn how agents fail against it. Find a published write-up of an attack that exploits that risk, then reproduce it on a target you are allowed to test, such as Gray Swan Arena’s weekly Proving Ground challenges. Write down what happened with the evidence first and your conclusion after it, including what the affected system stood to lose. That write-up is a work sample you can show an interviewer. When you interview for a seat that reviews agents, ask whether the agents record the evidence behind their verdicts and who goes over the cases they escalate.
If you have been in security for years, agent security is new to you as well, which makes this a good time to learn it alongside the people you supervise. If you lead a team that is adding agents, staff the review of their work as a training seat. Give it a short list of checks for each type of alert. Name an experienced practitioner to go over the escalations with each new reviewer every week. Keep the new reviewer’s calls in the record next to the agent’s so the review has something to correct. That one decision turns a review queue into the place where your team’s first agent security cadre learns the work.
A note on where I stand: I spent years leading security teams inside large financial institutions. I also sit on the advisory board of a company building autonomous security tools, which is the category whose work this issue describes. Nothing non-public, from anywhere I have worked, goes into what I write here.
Next: if you are trying to break in right now, the last issue of the season is for you.
