Skip to content

AgentGuard v1.0.0

Choose a tag to compare

View all tags
@nizba06 nizba06 released this 15 Jul 17:28
· 9 commits to master since this release
v1.0.0
f4667ef

inter-agent-guard v1.0.0 — Release Notes

Released: 2026-07-15
PyPI: inter-agent-guard · Import: agentguard
Repo: https://github.com/nizba06/agentguard

Highlights

  • Holdout gates cleared (20% Anthropic carve-out, 160 adversarial + 1,000 benign):
    • Detection 99.4%, FPR 0.0%
    • ML layer catches ~95% of content attacks; rules catch the rest of detected cases
  • Default ONNX artifact is dynamic INT8 (~164 MB, SHA-256 pinned)
  • Ship tooling: scripts/check_v1_gates.py, scripts/ship_v1.ps1, run_benchmark_evaluation.ps1 -Holdout
  • Deterministic layers (capability / impersonation / trust) unchanged and production-ready

Install

pip install "inter-agent-guard==1.0.0"
# Download risk_scorer.onnx + model.sha256 from GitHub Releases into agentguard/models/
python scripts/verify_model.py   # from a checkout, or use install_model helpers
from agentguard import AgentGuard, CapabilityManifest

guard = AgentGuard(
    risk_threshold=0.85,
    require_ml_model=True,
    task_objective="Analyse Q3 competitor pricing",
)

ML model (required for enforce-mode scoring)

Artifact Notes
risk_scorer.onnx Dynamic INT8 (~164 MB); SHA-256 1f758299994793a033153c27a033eb998c5e37090d51da3ce0af1c4807fe6894
tokenizer.json / tokenizer_config.json Bundled in the wheel under agentguard/models/

Latency SLA (honest)

CPU ONNX P95 on holdout is ~3 s, not the original 15 ms design target. For high-QPS production:

  • Rules-only / capability / trust (no ONNX)
  • GPU ONNX Execution Provider
  • Async / monitor modes — see docs/source/latency.md

v1.0 explicitly documents this CPU exception; do not assume sub-15 ms on CPU with ML enabled.

Benchmark authority

Holdout (benchmarks/results/holdout_report.md) is the ship gate. Full-corpus scores can overstate quality when the scorer trained on that JSONL.

Dataset: https://huggingface.co/datasets/Nizba/agentguard-benchmark-v1

Verify gates locally

py -3.12 scripts/verify_model.py
.\scripts\run_benchmark_evaluation.ps1 -Holdout -RequireModel
py -3.12 scripts/check_v1_gates.py --allow-cpu-latency
Assets 4
Loading