atrosa

News

Fraud needs detection engineers

· John

TabaPay, a payments processor in Palo Alto, listed a job this year that most fraud teams have never written down. The title was Fraud & Risk Engineer. The holder would own every blocking rule in the risk tool, ship each rule as a package with its decision logic, assumptions, test cases and acceptance criteria, run it in shadow mode before it touched a payment, and "apply engineering discipline to rule authorship: versioning, testing, modularity, documentation". The posted pay was $125,000.

Security has had a name for that job since about 2020. The name is detection engineer. Fraud has the work and does not have the name, and the gap between the two is the subject of this post.

How security got the role

The practice came first. In December 2017 Palantir's incident response team published its Alerting and Detection Strategy framework. Every alert lived in a GitHub repository with a written goal, its blind spots, its known false positives and a validation step the framework describes as "similar to a unit test". The people writing those alerts were still called IR engineers.

The title arrived about two years later. By February 2020 a security vendor could publish a post called "So, You Want to Be a Detection Engineer?" and expect readers to know what it meant. In 2021 a SANS paper described detection engineering as "becoming a common term" and "still a maturing concept".

The institutions came in 2022, built by practitioners. DEATHCon, a hands-on conference for detection engineering and threat hunting, was founded that year. Detection Engineering Weekly sent its first issue on 30 November 2022, from the person who ran detection engineering at Datadog.

The surveys show the rest. In November 2023 SANS reported that detection work was still spread across SOC analysts, security engineers and threat hunters rather than held by a standalone role. By February 2025 the same survey found that 60% of organisations had a dedicated detection engineering team, about 80% funded the function, and 41% could not find people to hire. The most common salary band was $101,000 to $150,000. Eight years from a blog post to a profession.

What a detection engineer does

Strip the security vocabulary and the job is five steps. Form a hypothesis about a behaviour worth catching. Write it as code, in a repository, with the reasoning beside it. Test it against history before it runs on anything live. Measure what it fires on, then tune it or retire it. Keep the version history, so anyone can see what changed and why.

Security arrived at each step by losing track of what was running and why, somewhere past the hundredth rule. Fraud teams are arriving at the same place on rule builders, mostly without the vocabulary.

Fraud has the work and not the role

Read the fraud postings open this autumn and the detection engineer's job is there, cut into three.

The fraud strategy analyst, or payments risk analyst, tunes rules in a vendor's console and reports their precision and recall. Coinbase's current posting asks for exactly that, along with the "transition from heuristic rules to model-driven" controls. Intuit's asks the analyst to read the rule engine's decision logs, reconstruct the attack and recommend a rule change. These are the people who write detections. They write them in a form, and the form has no version history, no test suite and no diff.

The fraud data scientist builds the models. The investigator works the cases. Nobody owns the rule as a piece of engineering. The hypothesis lives in someone's head, the rule lives in a vendor's database, and the evidence that it ever worked lives in a dashboard.

Three things kept fraud from building the role. Fraud teams sat in operations and compliance, where an engineer is an unusual hire. Vendors built no-code rule builders for those teams, so the detection had no code surface to engineer. And the rules decide money movement, so a bad change costs more than a missed alert, which argued for keeping changes rare and small.

The TabaPay posting is what it looks like when a payments company stops accepting that. It is one posting. There will be more.

The role will not be built the way security built it

Security built detection engineering with people. One engineer wrote one rule at a time, another reviewed it, and the profession took eight years to form. That is why 41% of security teams cannot hire for it today. The practice scaled and the people did not.

Fraud gets to build the role in 2026, and in 2026 the cheapest detection engineer is an agent. The five steps survive intact. What changes is who does each one.

An agent forms hypotheses by the hundred against a company's own exports, and the exports are the only input it needs. It writes each one as code, runs it in a sandbox over the full history, and keeps the ones that find something. The documentation Palantir asked for in 2017, the goal, the blind spots, the false positives, falls out of the test as a by-product. On one remittance operator's exports, with no labelled cases, 51 runs produced 2,037 rule versions, of which 12 were kept. A person wrote none of them. A person reviewed every one before it ran on new data.

That review is the job that remains, and it is a better job than the one it replaces. The reviewer reads a diff instead of clicking a threshold. They see the hypothesis written down, the replay result, and the accounts the rule would have caught last quarter. They sign off, or they send it back. Security calls this person a detection engineer. TabaPay called them a risk engineer. The name matters less than the job, and the job is now one reviewer for a whole rule base.

It will not be pure detection engineering

Fraud borrows the discipline and changes two parts of it.

The test is different. A security team validates a detection by replaying the attack, from a red team or an emulation tool. A fraud team validates against history: the confirmed cases, the reversals, the chargebacks, the one account the analyst already knew about. The oracle is the ledger. A detection that cannot be replayed over last quarter's ledger has not been tested.

The sign-off is different. A security detection that fires wrongly costs an analyst an hour. A fraud rule that fires wrongly holds a customer's money. So the rule runs in shadow mode first, and a named person, usually the MLRO or the head of fraud, owns the decision to let it block. The agent writes the rule. The code makes the call. A person signs the code.

Those two changes are where ATROSA plays. The agent does the authoring and the testing, in Python, in the customer's repository, against the customer's own exports. The customer's team does the review and the sign-off. When an examiner asks why an account was flagged, the answer is the rule that fired, which they can read and run again.

What to do next

If you run a fraud team you can start without us. Export your rules from the vendor console, even as JSON, and put them in a repository. Write the hypothesis for each one in a sentence above it. Before the next change, replay the current rule and the proposed one over the last ninety days and look at the difference. Those three steps are most of detection engineering, and no rule builder gives you any of them.

The rest is the agent doing the same thing at a scale a team never could. The open-source detections are on GitHub under the MIT licence. The hosted version runs your first export at no charge.

All news