Blog

Security reviews for research systems: start early, stay small

· 7 min read · 75bit

Security reviews have a reputation for arriving at the end of a project and finding things that are expensive to change. Research systems are especially prone to this, because they begin life on whatever was convenient: shared credentials, broad access to source data, notebooks reachable from anywhere on the network.

The cost is real, and it moves

IBM's annual Cost of a Data Breach report, based on research by the Ponemon Institute, tracks the average global cost of a breach. It was USD 3.86 million in the 2020 report, 4.35 million in 2022, 4.45 million in 2023 and 4.88 million in 2024. It fell to 4.44 million in 2025, which IBM links to faster detection and containment, then reached a record 4.99 million in the 2026 report.

Average global cost of a data breach

USD millions, by report year (the 2021 report is omitted here)

  • 3.862020
  • 4.352022
  • 4.452023
  • 4.882024
  • 4.442025
  • 4.992026
IBM, Cost of a Data Breach Report, 2020 to 2026 editions. Global averages across surveyed organisations.

Averages hide a wide spread, and a research platform is not an average organisation. The useful reading is the direction: the 2025 dip came from detecting and containing problems faster, which is exactly the kind of gain a small early review buys.

Start with data and access

Before reading a line of code, answer three questions. Most serious problems in research systems come from the answers.

QuestionWhat it usually uncovers
What data does this touch, and how sensitive is it?Personal or contractual data in a place it should not be
Who and what can reach it?Access granted once and never revisited; shared credentials
Where does it go afterwards?Outputs copied somewhere unmanaged

Keep the first pass lightweight

A first review should take an hour, not a quarter. Walk the data flow on a whiteboard, list the secrets and where they live, and note anything that depends on a person remembering to do something. Fix the cheap items at once and write the rest down with owners.

  • Replace shared credentials with per-person or per-service identities.
  • Move secrets out of notebooks and repositories into a managed store.
  • Give each environment its own access, so experiments cannot read production data by default.
  • Log access to sensitive data, so a question later has an answer.

Revisit at each stage

Repeat the same short review at each promotion: prototype, pilot, production. The cost stays low because the system is understood, and the findings arrive while there is still time to act on them. Security that grows alongside the work rarely needs a dramatic retrofit.

Sources