Jobs / Security
Can agents do security work?
Find, exploit, and fix vulnerabilities; monitor and defend systems.
- US jobs
- 190,650 (0.1% of employment)
- Typical wage
- $129K a year (employment-weighted mean of occupation medians)
- Occupations
- 4 in O*NET 31.0
- Evidence
- 7 direct benchmarks, 2 also relevant, 1 context only
The answer today
Nearly all tasksUp from Not yet on 10 Oct 2025.
Rule: Agents complete 90% or more of the tasks (or projects) on two direct benchmarks from independent institutions.
The evidence that sets this step
Next step: Some projects needs 20% or more of whole projects on two direct benchmarks from independent institutions.
How much of the work is measured
35% of the work in this family is measured by direct benchmarks. Some of the work is measured (20% to 50%). On 10 Oct 2025: 35%.
Work = the O*NET 31.0 work activities of the family's occupations, weighted by how important each task is and by US employment. A benchmark measures an activity when its tasks make the agent do it and its grading checks it. Method.
| Work activity (O*NET) | Share of the work | Direct benchmark |
|---|---|---|
| Implement security measures for computer or information systems | 25.8% | Measured |
| Develop operational or technical procedures or standards | 10.8% | Not measured |
| Maintain operational records | 9.8% | Not measured |
| Train others to use equipment or products | 9.2% | Not measured |
| Test performance of computer or information systems | 9.1% | Measured |
| Maintain current knowledge in area of expertise | 8.8% | Not measured |
| Monitor operation of computer or information technologies | 8.7% | Not measured |
| Direct organizational operations, activities, or procedures | 8.4% | Not measured |
| Resolve computer problems | 4.7% | Not measured |
| Coordinate with others to resolve problems | 4.7% | Not measured |
ARTEMIS live penetration test (agents vs. professionals) is a field study: it measures real deployments. That adds to the work measured, but a field study gives no task success rate, so it cannot set a step.
Our reading
Agents complete nearly all tasks on two direct benchmarks, but both are narrow. Anthropic reports that Claude Mythos Preview solved 100% of a 35-task Cybench subset of capture-the-flag tasks. On CyberGym, a team entry reproduces 99.2% of 1,507 known vulnerabilities; teams run and submit their own results. Real web targets are harder: on CVE-Bench the best agent exploits 40% of 40 web applications, given a vulnerability description. No direct benchmark measures developing security procedures or maintaining operational records. In one live penetration test, the ARTEMIS scaffold scored above the median of ten professionals; that field study gives no completion rate and does not set the step.
Evidence: Cybench, CyberGym, CVE-Bench, ARTEMIS live penetration test (agents vs. professionals) by Carnegie Mellon University, Gray Swan AI, Stanford University, UC Berkeley, and UIUC. Written 10 Oct 2026.
Benchmarks for this work
Direct evidence
Work benchmarks mapped primary to this family. They set the step when they report how many tasks agents complete.
- 22Claude Mythos Preview (early checkpoint) · 13 Apr 2026Strong
- ARTEMIS live penetration test (agents vs. professionals)by Stanford University, Carnegie Mellon University and Gray Swan AI95.2ARTEMIS A2 (ensemble supervisor, Claude Sonnet 4 sub-agents) · 10 Dec 2025Strong
Also relevant
These benchmarks cover part of this family's work, or have no result yet. They do not set its step.
- 64.8%Claude Code + Opus 5.5 (max) · 6 Oct 2026Strong
Context only
These benchmarks span many occupations, or they test how agents behave. They give context but do not set the step.
Functions: Penetration testingLong-horizon autonomyVulnerability researchVulnerability patchingIncident responseCompliance operationsFinOpsTerminal operationsFeature developmentML engineering
What we do not know
- Benchmarks measure tasks, not whole jobs. A job is a bundle of tasks, judgment, and relationships.
- Capability is not adoption. A task that agents can do in a benchmark may still be done by people at work.
- Most results come from single runs. Reliability over many runs is often unknown.
Occupations in this family
Sorted by US employment (BLS OEWS May 2025). O*NET sub-occupations share the employment figure of their SOC group. Work measured = the share of the occupation's O*NET work that direct benchmarks test.
| Occupation | SOC | US jobs | Work measured |
|---|---|---|---|
| Digital Forensics Analysts | 15-1299.06 | (435K group) | 7% |
| Information Security Engineers | 15-1299.05 | (435K group) | 7% |
| Penetration Testers | 15-1299.04 | (435K group) | 7% |
| Information Security Analysts | 15-1212.00 | 191K | 35% |
How to cite
Credit the original work first: the benchmarks listed on this page, and their institutions.
Then, if you used this page:
Can Agents Work. "Can agents do security work?." https://canagentswork.com/jobs/security/ (accessed 2026-10-11). CC BY 4.0.@misc{caw-jobs-security,
title = {{Can agents do security work?}},
author = {{Can Agents Work}},
year = {2026},
howpublished = {\url{https://canagentswork.com/jobs/security/}},
note = {Accessed 2026-10-11. CC BY 4.0}
}