Penetration Testers
In short: today's best AI models could do about 60% of this job's working time (projected from METR's April 2026 measurement and Epoch's November 2026 index) if the work were set up for them, and about 96% by the end of 2028 if progress keeps its long-run pace. The people and hands-on work, about 3% of the time, stays with people.
Evaluate network system security by conducting simulated internal and external cyberattacks using adversary tools and techniques. Attempt to breach and exploit critical systems and gain access to sensitive information to assess system security.
$116,580 a year (for the broader BLS group this job sits in).
A typical 40-hour week
Not yet reviewed task by task- AI could take on24 h
- A person checks AI's work0 h
- Desk work AI can't do yet15 h
- Stays with people1 h
Penetration Testers, now: of a 40-hour week, AI could take on 24 hours, 0 hours become checking its work, 15 hours are desk work it can't do yet, and 1 hours stay with people.
The largest pieces: Document penetration test findings (about 2 h); Identify new threat tactics, techniques, or procedures used by cyber threat actors (about 2 h); Write audit reports to communicate technical and procedural findings and recommend solutions (about 2 h).
Estimates for a typical week of this job, if progress keeps its long-run pace, counting work the best models finish four times in five. Every task is below.
of the working time. Now is projected from METR's April 2026 measurement and Epoch's November 2026 index.
- 97%
- 1%
- 3%
of the work: Anthropic's measure of what people already bring to Claude at work (2025). It counts help with part of a task, so it can be higher than the share within reach.
On the chosen pace and reliability. Estimate.
Share of the working time within reach, now to 2030
Change assumptionsHide assumptionsLong-run pace, four in five
What each choice means: , , , , . Every pace starts from .
At this date and pace, models finish tasks up to 2.4 work days long four times in five.
How far this has come
With GPT-4, measured in March 2023, 0% of this job's working time was within reach at . With Claude 3.7 Sonnet, measured in February 2025, 0%. With Claude Mythos Preview (early), measured in April 2026, 8%. Today's estimate: 60%.
A backcast: today's task data with the best model METR had measured by each date. It shows how fast the models have moved, not what anyone forecast at the time.
Measured use over time
Of the Claude.ai conversations Anthropic sampled and tied to a work task, 0.027% went to this job's tasks in May 2026, the 292nd largest of the 447 jobs that showed up at all; 3 of its 22 tasks showed up. Within each of Anthropic's methods, never across a change: 0.051% to 0.027% (April to May 2026).
Each release
| Conversations | Share | Rank | Tasks seen |
|---|---|---|---|
| Released February 2025 | none | n/a | 0 of 22 |
| Released March 2025 | none | n/a | 0 of 22 |
| The week of Aug 4, 2025 (released September 2025) | none | n/a | 0 of 22 |
| The week of Nov 13, 2025 (released January 2026) | none | n/a | 0 of 22 |
| The week of Feb 5, 2026 (released March 2026) | none | n/a | 0 of 22 |
| April 2026 (released June 2026) | 0.051% | 224 of 434 | 4 of 22 |
| May 2026 (released June 2026) | 0.027% | 292 of 447 | 3 of 22 |
Measured, not estimated: Anthropic's samples of Claude.ai conversations, each tied by a model to an O*NET task, from the Anthropic Economic Index. It says where Claude's use goes, not how much of this job's working time it covers, so it is not on the same scale as the reach above. The releases differ in their sample, their classifier and the smallest share they report (the June 2026 file rounds each task to a hundredth of a percent, so a smaller task reads as none), which is why a change of method starts a new line. At least 96% of each older release's task-tied use matched this site's tasks, and all of the newest. Claude.ai only, so use through other tools and other companies' models is not in it. Every job ranked.
Robots and the physical work
Anthropic rates 1% of this job's working time as physical work. For 0% of the working time, a robot has been shown doing most of the task somewhere. 0.00 of 3.
- Not physical 99%
- No robot shown 1%
Anthropic's rating, not this site's estimate: Claude Opus 5, searching the web, rated each physical O*NET task by the kind of setting in which a robot it could cite (deployed, sold or demonstrated) does most of the task, weighted by its own estimate of each task's share of the time (What work can robots do?, September 30, 2026, data CC BY 4.0; every job ranked). A robot shown somewhere is not a robot in your building: cost, setup and safety decide that. The share within reach on this page still counts no body work. The robot behind each rated task is under the task below.
The tasks
Sorted by . 22 tasks from .
| Task | Kind | |||
|---|---|---|---|---|
| Evaluate vulnerability assessments of local computing environments, networks, infrastructures, or enclave boundaries.Done more than monthly by most who do it. | 7% | 6.7 h1.7 h to 1.7 work days | 69%Half by now | |
| Write audit reports to communicate technical and procedural findings and recommend solutions.Done more than monthly by most who do it. | 7% | 1 work day3.3 h to 2.5 work days | 69%Half by now | |
| Identify security system weaknesses, using penetration tests.Done more than yearly by most who do it. | 7% | 1.7 work days6.7 h to 2.5 work days | 67%Half by now | |
| Conduct network and security system audits, using established criteria.Done more than monthly by most who do it. | 7% | 1.7 work days6.7 h to 2.5 work days | 67%Half by now | |
| Test the security of systems by attempting to gain access to networks, Web-based applications, or computers.Done more than yearly by most who do it. | 7% | 1.7 work days6.7 h to 1 work week | 43%Half by Dec 2026 | |
| Develop and execute tests that simulate the techniques of known cyber threat actors.Done more than yearly by most who do it. | 7% | 1.7 work days6.7 h to 1 work week | 43%Half by Dec 2026 | |
| Develop infiltration tests that exploit device vulnerabilities.Done more than yearly by most who do it. | 7% | 1.7 work days3.3 h to 1 work week | 28%Half by Apr 2027 | |
| Develop security penetration testing processes, such as wireless, data networks, and telecommunication security tests.Done more than yearly by most who do it. | 7% | 2.5 work days6.7 h to 1.3 work weeks | 20%Half by May 2027 | |
| Investigate security incidents, using computer forensics, network forensics, root cause analysis, or malware analysis.Done more than monthly by most who do it. | 7% | 2.5 work days6.7 h to 1.7 work weeks | 18%Half by Jun 2027 | |
| Document penetration test findings.Done more than yearly by most who do it. | 6% | 3.3 h1.7 h to 6.7 h | 100%Half by now | |
| Identify new threat tactics, techniques, or procedures used by cyber threat actors.Done daily by most who do it. | 6% | 3.3 h1.7 h to 1.3 work days | 97%Half by now | |
| Gather cyber intelligence to identify vulnerabilities.Done more than weekly by most who do it. | 5% | 3.3 h1.7 h to 6.7 h | 100%Half by now |
Within reach, per task: the share of the task's time whose instances are short enough for a model to finish at the chosen date, pace and reliability. Half by: the date half of that time comes within reach.
Jobs that share skills with this one
From O*NET's related occupations. Within reach now and by the end of 2028, long-run pace, four in five.
- Information Security Engineers53% now, 97% by 2028
- Information Security Analysts52% now, 87% by 2028
- Software Quality Assurance Analysts and Testers66% now, 91% by 2028
- Computer Systems Analysts35% now, 89% by 2028
- Computer Systems Engineers/Architects45% now, 67% by 2028
- Software Developers48% now, 78% by 2028
Where this job works
The industry groups that employ the most of it (BLS, May 2025). Counts are for the broader BLS group this job sits in.
- Federal Executive Branch86,390 people
- Computer Systems Design and Related Services76,660 people
- Management of Companies and Enterprises25,080 people
- Computing Infrastructure Providers, Data Processing, Web Hosting, and Related Services16,520 people
- Management, Scientific, and Technical Consulting Services15,100 people
Where the most of these jobs are
The metro areas with the most people in it (BLS, May 2025). Counts are for the broader BLS group this job sits in.
Questions about this job
- How much of the work of Penetration Testers is within reach of AI now?
- An estimated 60% of the working time is within reach of AI models now and 96% by the end of 2028, on METR's long-run pace at four in five (92% to 97% across the three paces the site shows). Within reach means a model could finish work that long at that reliability. It is not a forecast of jobs lost or of savings.
- Which tasks of Penetration Testers come within reach first?
- Weighted by the time each takes, the tasks with the most within reach now (projected from METR's April 2026 measurement and Epoch's November 2026 index) are: Document penetration test findings (100% of the task's time now); Identify new threat tactics, techniques, or procedures used by cyber threat actors (97% of the task's time now); and Write audit reports to communicate technical and procedural findings and recommend solutions (69% of the task's time now, 100% by the end of 2028).
- Which tasks of Penetration Testers stay with people?
- Tasks that are people work or body work are not counted as within reach at any date. The largest: Assess the physical security of servers, systems, or network devices to identify vulnerability to temperature, vandalism, or natural disasters; and Discuss security solutions with information technology teams or management. Together such tasks are about 3% of the working time.
- Do people already use AI for the work of Penetration Testers?
- In the Claude.ai conversations Anthropic sampled for May 2026 and tied to a work task, 0.027% went to this job's tasks, the 292nd largest share of the 447 jobs that showed up; 3 of its 22 tasks appeared. Anthropic's labor market measure puts the share of the job's tasks seen in real Claude use at work at 31%. These are measures of use, not of how much of the work AI could finish, and they cover Claude only.
- Can robots do the physical work of Penetration Testers?
- Anthropic's robot exposure rating (September 2026) finds 1% of the working time is physical work, and that a robot has been shown doing most of the task, somewhere, for 0% of the working time. Its index is 0.00 of 3. A robot shown somewhere is not one in every workplace; this site's own estimates still count no physical work as within reach.
- How is this estimated?
- From the 22 tasks O*NET lists for the occupation, each with an estimated length checked against GDPval and a share of the working time, each classed as desk, people or body work, and METR's measurements of how long a task the best AI models can finish, carried forward to the date from the latest, Claude Mythos Preview (early) in April 2026. Data checked November 10, 2026.
This job in your state or metro area
Embed these figures on your site
A small card with the share of working time within reach now and at the end of 2027, 2028 and 2030, the date and a link back. Free under CC BY 4.0; the card carries the credit.
Your own version of this job
The page above is the occupation. Your job is your title and what you actually do: give both, and they are read against these tasks in your browser, nothing sent anywhere. The headline is free: how much of your working time is within reach now and by the end of 2028, and how many of the tasks your description names.
This occupation has not been reviewed task by task yet, so the full personal report is not offered for it; 402 of 923 occupations are reviewed so far, largest first.