Refolk
10 min readMarkdown

133 AI Security Engineers vs 1,242 Pen Testers: Which Bullets Transfer

CyberSN counts AI security postings by the work described. Here's how pen testers can reframe authorization testing honestly, and which LLM claims need proof.

If you test applications for a living, the 2026 cybersecurity job numbers look grim. CyberSN's Cybersecurity Jobs Report has penetration tester postings down 45.9% and AI security engineer postings up 268.3%. The part that matters for your resume is in the method: CyberSN sorts postings by the work they describe, not by the title on them. So your bullets decide which bucket you land in, as long as you've actually done the work they describe.

What the CyberSN report actually counted

US cybersecurity postings fell 26.4% to 235,786 from January through September 2026, while AI security engineer postings rose to 2,409 over the same period. These are postings, not hires, and the report does not claim AI caused the losses.

The headline figures, same nine months year over year:

  • All security postings: 235,786, down from 320,210.
  • AI Security Engineer: 2,409, up 268.3%.
  • Penetration Tester: 3,063, down 45.9%. They fell another 26.6% from Q2 to Q3 2026.
  • Application Security Engineer: up 9.9% from Q2 to Q3, the same quarter pen testing fell.
  • CISO: 2,832, down from 4,632 (a 38.9% drop).

Some of that drop is a return to normal after a spike. 2025 was an unusually strong year, and compared with 2024 the overall gap shrinks to about 4%. CISO postings in 2026 (2,832) nearly match 2024's 2,808. Deidre Diamond, CEO of CyberSN, has said the swings in leadership hiring reflect disruption in the C-suite more than a falling need for security leaders. The report also states that it doesn't determine why postings declined, and it doesn't establish that AI adoption caused cuts in other security roles.

268.3%
Growth in US AI Security Engineer postings, Jan to Sep 2026

From roughly 654 postings to 2,409, per CyberSN. Fast growth, small base.

Why 268% growth is a small market with a big edge

The growth in AI security postings doesn't replace the lost pen testing work. AI security engineer postings make up about 1% of all US security postings (2,409 of 235,786). The advantage is competition: very few people hold the title yet.

From CyberSN's own numbers, AI security postings gained about 1,755 from an implied 2025 base of about 654, while pen tester postings lost about 2,600 from an implied base of about 5,660. If you read "268%" as "the market moved to AI," you'll overestimate how many of these jobs exist.

Supply is where it gets interesting. In Refolk's index of professional profiles, 1,242 US profiles hold a pen tester, pentester, or offensive security engineer title. Only 133 hold an AI security engineer, AI red team, LLM security, or AI security researcher title.

Role familyUS profiles (Refolk's index)Postings, Jan to Sep 2026 (CyberSN)YoY changePostings per profile (derived)
Pen tester / offensive security1,2423,063−45.9%2.5
AI security engineer / red team / researcher1332,409+268.3%18.1
Headline mentions prompt injection or LLM red teaming10n/an/an/a

Pen tester profiles outnumber AI security profiles 9.3 to 1, but pen tester postings outnumber AI security postings only 1.27 to 1. That works out to roughly 7 times as many postings per person on the AI security side.

Read the ratio as a direction, not a precise number. It compares postings classified by work against profiles classified by title. The direction is still clear: the gap exists because the title is new, not because the role is big. A penetration tester career change into AI security is a bet on scarcity, and scarcity rewards people who can prove the work.

The method is the hook: your bullets pick your bucket

CyberSN classifies postings by the work described, so the way you describe your work decides which bucket you match. Describe authorization testing as defensive design work and you line up with the buckets that are growing. Describe it as a vulnerability count and you line up with the one that's shrinking.

The report says it directly: "demand is counted by the work being asked for, not by the title on the listing." Two details in the method change how you should write:

  • AI Security Engineer is an overlapping series. CyberSN shows it for context but doesn't add it to the total. One posting can count as both AppSec work and AI security work, so AppSec phrasing doesn't cost you AI security relevance.
  • AI security is filed under defense, not offense. CyberSN puts AI security engineer roles in the "defense" function. A bullet that only says "found [n] critical vulns" matches the offense bucket that fell 26.6% in a single quarter. A bullet about scoping permissions, writing remediation, and designing least-privilege access matches the defense classification.

Here's what that looks like in practice:

  • Before: "Performed web app penetration tests and identified [n] high-severity findings."
  • After: "Tested authorization boundaries across [n] customer-facing APIs; wrote least-privilege remediation that engineering shipped in [timeframe]."

Both bullets describe the same engagement. The second one describes the defensive half that you also did. Doing that for every posting you apply to is the tedious part, and it's the exact work Refolk takes off you. Paste the posting, and I rewrite your own history toward the language it asks for, without inventing a project you didn't do.

Which pen test work honestly maps to AI security requirements

Access-control and authorization work maps most cleanly, because OWASP's Excessive Agency risk is an access-control problem with a language model in the middle. Excessive agency is when an LLM-driven agent can take actions beyond what its task needs, because it has too many tools, too much access, or too little oversight.

OWASP's mapping for LLM06:2025 (Excessive Agency) lists three root causes:

  1. Excessive functionality: the agent has tools it doesn't need.
  2. Excessive permissions: those tools run with broader access than the task requires.
  3. Excessive autonomy: high-impact actions happen without a human check.

The fixes OWASP recommends are fewer tools, narrower scopes, and complete mediation. An access-control reviewer already does all three. One caution: always cite the version year with the ID. LLM03 meant Training Data Poisoning in 2023/24, Supply Chain in 2025, and Excessive Agency in 2026. Writing "OWASP LLM03" with no year tells an interviewer you skimmed a blog post.

Work you've doneAI security requirement it maps toHonest framing
IDOR/BOLA testingAgent permission boundaries (excessive permissions)"Tested authorization boundaries"
Scoping service accounts to least privilegeTool and credential scoping for agents"Designed least-privilege access for automated services"
Abuse-case threat modelingThreat modeling AI features and tool misuse"Built abuse cases for multi-step workflows"
Writing remediation guidanceMitigation design (narrow scopes, mediation)"Wrote remediation that removed unneeded access paths"

IDOR and BOLA (insecure direct object reference and broken object-level authorization) are flaws where a user can reach records they shouldn't, usually by changing an ID. Testing for them thoroughly is real evidence that you think about authorization boundaries. That's exactly what an agent with a CRM tool needs someone to think about.

The claims you cannot make yet

You can't write "prompt injection testing experience" on a resume until you've tested an actual LLM application, worked through a hands-on lab, or published a write-up. IDOR testing earns "tests authorization boundaries." It doesn't earn that.

Prompt injection is an attack where text the model reads overrides its instructions. In the indirect version, the malicious text arrives through data the model processes, such as a document or a form field, not from the person chatting with it.

10
US profiles whose headline mentions prompt injection or LLM red teaming

In Refolk's index. They include a generative AI red teamer at Meta and people at Anthropic and Walmart.

That scarcity works both ways. A real prompt injection testing line on your resume stands out against a field of 10. A fake one is easy to catch, because the interviewer will ask how you'd find an injection path through a lead form, or what permissions you'd strip from the agent behind it.

What each claim needs behind it:

  • "Tests authorization boundaries": IDOR/BOLA or access-control testing you've done. Supported.
  • "Threat-modeled AI features": an actual threat model of an LLM feature, even an internal one.
  • "Prompt injection testing experience": a tested LLM app, a completed lab, or a CVE write-up.
  • "AI red teaming": engagements against a model or agent, with findings you can walk through.

Refolk's fit score helps here because it scores against what your history actually supports. A prompt injection requirement you haven't touched shows up as a gap you can go close, not a bullet you'll have to defend live.

A real prompt injection bullet stands out against a field of 10. A fake one gets caught in the first follow-up question.

How to become an AI security engineer from pen testing

The fastest honest route is to pick up one or two pieces of real LLM work, then rewrite the bullets you already have so they show the defensive side. Here's the order that works:

  1. Work a hands-on lab. Pwned Labs' AI Systems Attack and Defense bootcamp covers prompt injection, tool abuse, and MCP attacks. Completing it supports a lab-scoped bullet.
  2. Reproduce and write up a public incident. ForcedLeak (CVSS 9.4, 2025, reported by Noma) was an indirect prompt injection in Salesforce Agentforce, delivered through a Web-to-Lead field, that led to CRM data exfiltration. It combines LLM01 with LLM06. EchoLeak (CVE-2025-32711) was a zero-click prompt injection in Microsoft 365 Copilot. Write up how the chain works and which permission would have broken it. Don't claim the find.
  3. Volunteer to test your employer's LLM feature. Scope it like any engagement: access, tools, data sources. That gives you real work that names the system.
  4. Then rewrite. Put AppSec and access-control language first, and add the LLM bullet that the lab or the engagement now supports.

For background reading, Kai Greshake's "Inject My PDF: Prompt Injection for your Resume" is cited in OWASP's LLM Top 10 references. Read it to understand indirect injection. Don't try it on your own resume. Nothing tells a hiring manager you misread the job faster.

Who already holds the title, and how they get found

In Refolk's index, 16 of a 25-profile sample of AI security people carry the exact title "AI Security Engineer," at employers including Cisco, eBay, 7AI, Hack The Box, and Wraithwatch. In a 25-profile sample of pen testers, Meta employs 7, and one of them already uses the title "Offensive AI Security Engineer."

Since that exact title is the one most holders in the sample use, use it when your work supports it. Don't expect an easy entry point, though. CyberSN identified only 9.5% of all security postings as entry level. That makes this a lateral move for experienced testers, not a first job.

There's a catch in how this market finds people. Only 133 US profiles carry an AI security title, so anyone searching by that title sees a short list. A pen tester with real LLM lab work and strong authorization bullets doesn't show up on it. They end up applying cold into the offense bucket that dropped 26.6% in a quarter. And if you're employed on a security team, putting "open to AI security roles" on a public profile tells your current manager exactly what you're doing.

FAQ

Is penetration testing dying because of AI?

The data doesn't show that. Pen tester postings fell 45.9% to 3,063, but 2025 was an unusually strong year, and the overall security market is only about 4% below 2024. CyberSN says its report doesn't determine why postings declined or establish that AI caused cuts. Treat the drop as partly a reversal of a spike, plus a shift toward defense-side and AppSec work, where Application Security Engineer postings rose 9.9% in the same quarter pen testing fell 26.6%.

Can I list prompt injection testing experience on my resume if I've only done IDOR testing?

No. IDOR and BOLA testing honestly supports "tests authorization boundaries." Prompt injection testing experience needs an actual tested LLM application, a completed lab such as Pwned Labs' AI Systems Attack and Defense bootcamp, or a published write-up. Only 10 US profiles in Refolk's index mention prompt injection in their headline, so interviewers will dig into that claim.

How big is the AI security engineer market, really?

Small. CyberSN counted 2,409 AI security engineer postings from January through September 2026, about 1% of all US security postings, and only 9.5% of security postings overall were entry level. The advantage is low competition, not volume: only 133 US profiles in Refolk's index hold an AI security title, compared with 1,242 holding pen tester titles.

The data behind this

Put this to work

Paste your career in once. Every application after that is written for you.

Drop a resume or a LinkedIn URL. I rank the live openings against it, rewrite the resume and write a cover letter for the best of them, and fill in the employer's form when you press the button. You read, you decide what goes out.

  1. 01Drop your resume

    A PDF or a LinkedIn URL. About a minute, once.

  2. 02I rank the openings

    Every weekday morning, the live catalog scored against your history. Up to 20 worth your time, not two hundred links.

  3. 03Each one is written up

    Resume rewritten for the posting, a cover letter, a fit score. Press send, or let me fill in the form.

  • New matches ranked and written before you are up.
  • Every bullet stays inside what your history supports. Nothing invented.
  • Queued, submitted, interviewing, offer: one screen, not a spreadsheet.

Free, no card. Up to 25 applications sent a week, and nothing is sent until you say so.

Keep reading