Cybersecurity Engineering + AI: The 2026 Career Guide
Updated August 17, 2026·10 min read
TL;DR
In one month — March to April 2026 — AI/ML requirements in security JDs went from 8% to 19%. The gap is in protecting AI systems from a new class of attacks, not using AI for threat detection. Security engineers already own the hard part — threat modeling, trust boundaries, adversarial mindset. The AI knowledge adds on top.
In one month — March to April 2026 — AI/ML requirements in security-engineering JDs went from 8% to 19%, a 2.5× jump in a single month. The demand is not for using AI to detect threats — it is for protecting AI systems from a new class of attacks. Security engineers already own the hard part — threat modeling, trust boundaries, an adversarial mindset — and the AI knowledge adds on top.
Key facts
- In Dexity's April 2026 analysis of 273 US security-engineering JDs, 19% required AI or ML skills — up from 8% one month earlier in March 2026.
- Dexity's JD data shows demand is asymmetric: of 273 security JDs, 19% asked for AI skills, while of 139 AI engineering JDs only 0.7% asked for security skills.
- 83% of AI/ML security demand in Dexity's April 2026 analysis sits at senior level and above.
- Disclosed senior AI security salaries in Dexity's US dataset run $180K–$275K, with lead/manager roles reaching $300K.
- The Generative & Agentic AI cluster in security JDs went from zero in March 2026 to 24 JDs in April 2026 — a single month.
- In Dexity's April 2026 dataset, the AI/ML Security Challenges cluster — naming prompt injection, model poisoning, and unsafe autonomy — appeared in 9% of all security JDs (24 JDs).
What does "AI security" actually mean?
The term covers two completely different disciplines. Getting this wrong wastes months.
"AI for Security" — using machine learning to improve threat detection, automate SOC workflows, classify malware, speed up incident triage. The credential and institution space owns it. No real gap.
"Security for AI" — protecting AI systems themselves: LLMs, RAG pipelines, agent workflows, ML model pipelines. New attack surfaces that didn't exist three years ago. This is the gap.
The new attack surfaces:
Prompt injection — an attacker inserts malicious instructions into a user input or a retrieved document (in a RAG pipeline), causing the model to act outside its intended behaviour. Can exfiltrate data, bypass guardrails, or make agents take unintended actions.
Model poisoning — corrupting training or fine-tuning data so the model behaves differently under specific trigger conditions. Hard to detect after the fact; devastating if the model is used in a high-trust context.
Agent vulnerabilities — multi-agent systems introduce attack paths that don't exist in single-model deployments: tool call injection, memory poisoning, inter-agent trust assumptions.
Supply chain attacks — compromised model weights, malicious dependencies in the ML stack, poisoned datasets from third-party sources. The ML ecosystem has an attack surface most security teams aren't yet covering.
None of these map onto firewall rules, endpoint detection, or SIEM playbooks. Threat modeling still applies — the mental model shifts.
Where does an AI security engineer sit in the ecosystem?
You're adding a specialisation on top of what you already have — not replacing the fundamentals.
| Role | Primary focus | Key difference |
|---|---|---|
| SOC Analyst | Monitor and respond to incidents | Reactive, traditional threat detection. AI Security Eng is proactive and AI-system-specific. |
| Penetration Tester | Attack traditional systems — networks, apps, APIs | AI Security Eng extends pen testing to LLMs, RAG pipelines, agent systems. |
| DevSecOps Engineer | Integrate security into the dev pipeline | Shares the integration mindset but doesn't yet cover AI-specific attack surfaces. |
| Security Architect | Design security posture at system level | AI Security Eng applies architectural thinking specifically to AI systems and their new failure modes. |
If you understand threat modeling and LLM attack surfaces, you can do work that none of these adjacent roles currently covers.
What does the April 2026 JD data show?
273 LinkedIn security engineering JDs · US market · April 2026 · 20% disclosed salary
The direction of demand
AI skill clusters being hired for
| Cluster | % of all JDs | What employers actually want |
|---|---|---|
| AI/ML Security Challenges | 9% (24 JDs) | Prompt injection, model poisoning, unsafe autonomy, tool misuse |
| Generative + Agentic AI | 9% (24 JDs) | LLM-assisted threat hunting, agentic AI solutions |
| AI/ML Dev + Integration | 5.5% (15 JDs) | AI frameworks, agentic architectures, model fine-tuning |
| AI/ML Security Tools | 3% (7 JDs) | AI-driven SIEM analytics, ML-based detection |
The "AI/ML Security Challenges" cluster naming specific attack surfaces in JDs confirms this is no longer theoretical demand.
Seniority distribution
| Level | % of AI/ML roles |
|---|---|
| Senior | 52% |
| Lead / Manager | 31% |
| Mid | 13% |
| Entry | 2% |
83% of AI/ML security demand is at senior level and above. This is where the premium lives.
Companies actively hiring (April 2026)
Google, Uber, CoreWeave, Box, GitLab, Pinterest, Plaid, Snap, Nordstrom, 1Password, Ford, Royal Caribbean Group, Stellar Cyber, Vectra AI.
Salary ranges
| Level | Range |
|---|---|
| Entry | $95K–$158K |
| Mid | $130K–$190K |
| Senior | $180K–$275K |
| Lead / Manager | $190K–$300K |
Which security skills carry over to AI security?
Foundation — assumed across all roles: Network security, encryption, firewall management, incident response, Linux
Required by majority of JDs:
| Skill | % of JDs |
|---|---|
| Cloud security | 62% |
| Application security | 59% |
SIEM (Splunk, Microsoft Sentinel) |
55% |
| DevSecOps practices | 40% |
| Zero Trust | 35% |
Programming: Python is the baseline. Go (11%) and TypeScript (7%) appear for infrastructure-heavy roles.
Dexity Intel · free newsletter
Liking this? Get the next one in your inbox.
JD-backed career reads, AI market signals, and field-tested tool guides — a few times a month. No fluff, no spam.
What are the two paths into AI security?
Path 1 — Security engineer adding AI skills
What you already have: Threat modeling, trust boundary design, network security fundamentals, incident response, cloud security — and critically, an adversarial mindset. You already think about how systems fail under attack. That's the hard thing to develop.
What to add:
| Skill area | Why it matters |
|---|---|
| LLM fundamentals | Understand what prompt injection actually exploits — context windows, instruction following |
| Prompt injection | Manual attack techniques and automated testing — the most common employer ask |
| Agent attack surfaces | Tool call injection, memory poisoning, inter-agent trust — newest area, in 24 JDs |
| Model supply chain | Where ML models come from, dependency risks, training data integrity |
| Output control and guardrails | PII detection, hallucination monitoring, output validation in production |
Timeline: 6–10 weeks of hands-on work against realistic systems. You're not starting from zero — you're applying existing instincts to a new attack surface.
Path 2 — AI engineer moving into security
What you already have: LLM application architecture, RAG pipeline design, agent frameworks (LangChain, CrewAI, LlamaIndex), model evaluation patterns. You know the systems you'll be attacking and defending.
What to add — this takes longer:
| Skill area | What to expect |
|---|---|
| Threat modeling | STRIDE, attack trees, adversarial mindset as a discipline — 4–6 weeks to develop real fluency |
| Trust boundaries | Build it into architecture reviews — design habit, not a checklist |
| Secure code review | Different from correctness review; takes practice |
| IAM fundamentals | Foundational for any system with external access |
The honest take: AI engineers have a language advantage — you know the systems. The security intuition takes real time to build. The fastest path is working alongside experienced security engineers on real production systems.
What is AI security engineering not?
Not SOC work. SOC Analysts respond to threats. AI Security Engineers define and test defences for AI systems before those threats materialise.
Not traditional pen testing. Standard network and application pen testing doesn't cover AI-specific attack surfaces. Prompt injection, model poisoning, and agent vulnerabilities require different mental models and different tooling.
Not AI engineering. You're not building LLM features or RAG pipelines. You're attacking and securing them.
Not governance or compliance. NIST AI RMF, ISO 42001, and compliance frameworks are a different track — different hiring funnel, different skills. The AI security engineering track is technical, not advisory.
Why is AI red teaming a real opportunity?
AI red teaming — systematically attacking AI systems to find vulnerabilities before real attackers do — barely existed as a named discipline before 2023.
Why this is a real signal:
- OWASP LLM Top 10 was first published in 2023. Agentic AI red teaming as a named discipline is even newer.
- The Generative & Agentic AI cluster in security JDs went from zero in March to 24 JDs in April 2026 — one month.
- The people building hands-on credibility now are not late.
Agentic AI red teaming is the next inflection: Multi-agent systems are being deployed into production faster than any security practice is adapting. Tool call injection, memory poisoning, cross-agent trust assumptions — training content is nascent, tools are nascent, and employer ask is already climbing. The people who build credibility here in the next 12 months will be significantly ahead.
Build the Security-for-AI stack. Dexity's security courses map to exactly this gap — AI Security: Harden, Defend, Ship for securing AI systems end to end, AI Red Teaming for offensively testing models and agents, Autonomous Attacker Defense for defending against agentic attackers, and Adversarial Cloud Security Engineering for the cloud layer underneath. Each is project-based — you build hands-on credibility, not just a certificate.
Why is the window for AI security closing?
The field is three years old. OWASP LLM Top 10 was first published in 2023. You're not late to something that's been around for a decade.
Demand is moving from signal to requirement. 8% to 19% in one month. The pre-enforcement phase is ending faster than expected.
The premium is still pricing in. Salary bands are wide (senior: $180K–$275K) and the top hasn't compressed — the market hasn't agreed on what the full stack looks like yet.
Agentic AI is the next multiplier. Multi-agent systems are outrunning every security practice. JDs are already mentioning agentic red teaming (3%). The people building credibility in this specific area now are getting ahead of the next inflection, not just the current one.
Frequently asked questions
What is the difference between "AI for Security" and "Security for AI"?
"AI for Security" uses machine learning to improve threat detection, automate SOC workflows, and speed triage, and there is no real hiring gap there. "Security for AI" means protecting AI systems themselves, including LLMs, RAG pipelines, and agent workflows, and that is where the gap and the opportunity sit. Targeting the wrong one wastes months.
What new attack surfaces do AI security engineers defend against?
The main ones are prompt injection, model poisoning, agent vulnerabilities like tool call injection and memory poisoning, and supply chain attacks such as compromised model weights and poisoned datasets. None of these map onto firewall rules, endpoint detection, or SIEM playbooks, though threat modeling still applies.
Is AI security a strong career move in 2026?
The demand is asymmetric in your favor: of 273 security JDs, 19% asked for AI skills, while of 139 AI engineering JDs only 0.7% asked for security skills. The opportunity runs one direction, so competition from the other side is near-zero. AI/ML requirements in security JDs also jumped from 8% in March 2026 to 19% in April 2026.
How much do AI security engineers make?
Disclosed US ranges run from $95K to $158K at entry, $130K to $190K at mid, $180K to $275K at senior, and $190K to $300K at lead or manager level. The demand skews senior, with 83% of AI/ML security roles at senior level and above, which is where the premium lives.
How long does it take a security engineer to add AI skills?
About 6 to 10 weeks of hands-on work against realistic systems. You are not starting from zero; you already own the hard part, which is threat modeling, trust boundary design, and an adversarial mindset, and you are applying those instincts to a new attack surface.
Is this the same as SOC work or pen testing?
No. SOC analysts respond to threats after they surface, while AI security engineers define and test defences before threats materialise. Standard network and application pen testing also does not cover AI-specific attack surfaces like prompt injection, model poisoning, and agent vulnerabilities, which need different mental models and tooling.
Source: LinkedIn JD Analysis — March 2026 (100 JDs) + April 2026 (273 JDs, 20% disclosed salary); independently re-scanned Jul 2026 across public ATS boards (general security-engineering lens; AI/ML in the majority of JDs; bands ~$185–265K), archived at data/2026-07/ · YouTube and cross-channel research (March 2026) · JD dataset for this role · Dexity.com
Go from reading to doing · Dexity Course
AI Red Teaming
Security engineering job descriptions now ask for AI security skills — prompt injection exploitation, agentic attack surfaces, model supply chain risk — and most security engineers have never touched them. In 6 weeks you'll build a mapped, tested attack surface for any LLM system, exploit the full OWASP LLM Top 10 and the emerging Agentic AI Top 10 before attackers do, and produce the audit reports and ship/hold frameworks your organization will act on.
