Benchmark Methodology: 2,920 Adversarial Tool-Call Vectors

#1
by acnbartholomew - opened

About this Benchmark Dataset

acnbartholomew/btp-agent-redteam-evals provides a curated, ground-truth labeled evaluation bank for measuring the robustness of autonomous agent tool-calling boundaries.

Categories Included:

  • Shell Breakout Vectors: rm -rf, fork bombs, reverse shells (/dev/tcp, mkfifo), base64 execution, and disk partition wipes.
  • SQL Mutations: DROP TABLE, TRUNCATE, cascade drops, and blind SQL injection.
  • Cloud SSRF: AWS, GCP, and Alibaba cloud metadata endpoints (169.254.169.254, 100.100.100.200).
  • Secret Exfiltration: Extraction of sk-*, ghp_*, AWS credentials, and SSH private keys.
  • Benign Controls: Everyday developer CLI commands and read-only queries to measure false-positive rates.

Running Evaluations:

from datasets import load_dataset
from btp_guard import Guard

dataset = load_dataset("acnbartholomew/btp-agent-redteam-evals", split="train")
guard = Guard(strict=True)

# Run full evaluation

Feel free to open PRs or suggest new attack vectors!

Sign up or log in to comment