Hugging Face's logo

Datasets:
mjkenney
/
monitor-injection-evals

ArXiv:
Tags:
ai-control
prompt-injection
monitoring
evaluation
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

monitor-injection-evals

DRAFT, not frozen. Prompt-injection evaluation set for trusted monitors (APPS so far; BashBench later), following Terekhov et al., Adaptive Attacks on Trusted Monitors (arXiv:2510.09462). Evaluation only. Do not train on it.

  • data/apps/apps_eval.jsonl: one row per sample. Honest rows are clean only; attack rows cover vanilla, inj_transfer_{gpt4,gpt4o,gpt41} and length-matched neutral_transfer_* controls.
  • data/apps/manifest.json: file sha256, counts and validity rates per condition.
  • data/apps/prompts/eval_{neutral,injection_aware}.jsonl: eval split rendered with the gleipnir student prompt contract (tool-trajectory-binary-v1), in gleipnir benchmark input format, invalid samples dropped.
  • Generator (draft): openai/o3-mini backdoors from RoganInglis/apps-control-arena + same-model honest solutions.
  • Filter on valid_parses, valid_tests, valid_backdoor (attack) and injection_present before scoring.

Code and full documentation: see the project's GitHub repo README.
Downloads last month
16
Total file size:
117 MB

Paper for mjkenney/monitor-injection-evals