confideR

Protect confidential research data when using AI coding assistants

Overview

confideR is a companion R package to the paper "Protecting confidential research data when using AI coding assistants: A practical guide". It is designed for researchers in ecological, environmental, fisheries, agricultural, and social science settings who use AI tools for code development but cannot expose the underlying data.

It provides four core capabilities:

1. Confidential mode: Activate session-level protection that clears AI API keys, unloads AI packages, and blocks them from being loaded.

2. Session auditing: Detect your IDE (RStudio, Positron, VS Code), scan .Rprofile for AI auto-connections, check for AI API keys, scan the VS Code extension directory, and query the system process table for active AI agent processes.

3. Data fingerprinting: Extract a privacy-safe structural summary of your confidential dataset (column names, types, distributions) without exposing any raw values. Supports three obfuscation levels. Auto-detects confidential columns across fisheries, ecological, environmental, agricultural, and social science naming conventions.

4. Data simulation: Generate realistic simulated survey datasets with group effects, strata effects, seasonal patterns, observer coverage, and secondary observations — safe to use with any AI tool. Supports the develop on simulated data, run on real data workflow.

Installation and Example

# From GitHub
remotes::install_github("smithja16/confideR")

# Quick start
library(confideR)

# 1. Audit your session FIRST
audit_session()

# 2. Activate confidential mode (clears AI keys, blocks AI packages)
confidential_mode_on()

# 3. Simulate data for AI-assisted development
sim_data <- simulate_data(
  n_obs    = 2000,
  n_groups = 30,
  include_observer  = TRUE,
  include_secondary = TRUE )

# 4. Fingerprint a dataset - structural summary only, no raw values.
#    In real use, point this at YOUR confidential data instead:
#    fp_real <- fingerprint(real_data, mode = "summary", obfuscation = "partial")
fp_sim <- fingerprint(sim_data, mode = "summary", obfuscation = "partial")

# 5. Format fingerprint for copy-pasting into a browser AI chat
format_for_prompt(fp_sim)
# This summary is safe to paste into an AI tool (after checking)

# 6. The "fingerprint -> simulate" round trip: turn a fingerprint back into a
#    synthetic dataset that mirrors its structure and distributions.
#    (Runnable here on fp_sim; in real use you'd pass fp_real.)
sim_real <- simulate_from_fingerprint(fp_sim)

# 7. Write simulated data for your AI workspace
# write.csv(sim_real, "data/simulated_data.csv", row.names = FALSE)
# This file is safe to use with an AI tool (after checking)

# NOTE: with obfuscation, simulated columns carry alias names (ID_1, Coord_1,
# Var1, ...). Keep those aliases while developing with AI. Only on the secure
# machine, for the final run against real data, restore the original names:
# sim_real <- restore_names(sim_real, fp_real)
# Never paste restored names - or code/errors referencing them - into an AI tool.

# 8. Check if object contains raw data
library(mgcv)
M <- gam(log(response) ~ stratum + s(month, bs="cc") + s(year) +
           s(covariate_1) + s(group_id, bs="re"), data = sim_data)
contains_data_like(M)  #TRUE
contains_data_like(sim_data)  #TRUE
contains_data_like(fp_sim)  #FALSE

# 9. When done, turn off confidential mode
confidential_mode_off()

The "develop on fake, run on real" workflow with physical separation

  AI Machine (workshop)          Secure Machine (vault)
  ========================       ========================
  simulated_data.csv             real_data.csv
  + AI tools enabled             + NO AI tools
  + develop analysis code        + run final analysis
           |                              ^
           |--- transfer code only -------|
           |  (Git, paste, file transfer) |
           v                              |
  Code tested on fake data       Code runs on real data

Functions

Confidential mode

• confidential_mode_on() — activate protection
• confidential_mode_off() — deactivate and restore
• is_confidential_mode() — check status
• restore_api_keys() — recover keys from a session that didn't close cleanly (refused while confidential mode is active)
• api_key_status() — show which AI keys are live, backed up, or on disk

Auditing

• audit_session() — comprehensive audit report
• audit_ide() — detect IDE and AI features
• audit_rprofile() — scan .Rprofile for AI config
• audit_packages() — check for loaded AI packages and option residuals
• audit_env_keys() — check for AI API keys
• audit_processes() — scan system process table for AI agent processes
• confider_status() — one-line status summary

Fingerprinting

• fingerprint() — create structural summary
• is_fingerprint() — test whether an object is a fingerprint
• alias_map() — view alias-to-original mapping (local only)
• restore_names() — rename simulated alias columns back to originals (local only)
• format_for_prompt() — format for AI chat

Simulation

• simulate_data() — generate survey data from parameters
• simulate_from_fingerprint() — generate synthetic data from a fingerprint() (the "fingerprint → simulate" round trip)

Leak detection

• contains_data_like() — check if object contains raw data
• ensure_no_data_leakage() — error if raw data detected

Script scanning

• scan_script() — check .R/.Rmd/.qmd for data exposure patterns
• check_notebook_outputs() — check for rendered output in notebooks

Citation

If you use confideR in your research, please cite:

Smith J.A., Roff A., Brown C.J. (2026). Protecting confidential data when using AI
coding assistants: A practical guide. EcoEvoRxiv (preprint). DOI: https://doi.org/10.32942/X2FM30