Research

Research beyond the test

Papers we have published and fixes we have contributed to the open-source tools that secure AI.

Publications

Weighted Safety Refusal (WSR): A Reference-free, Severity-weighted, Dual-axis Metric for Evaluating LLM Refusal Behavior

SSRN preprint · June 2026

Flat refusal averages hide the failures that matter. WSR weights refusals by severity across two axes and is gaming-resistant by proof — in the pilot, Llama 3.3 70B scores 0.800 flat but 0.730 under WSR, and the gap is prompt injection.

Upstream contributions

4 merged, 8 open. Status recorded 8 Oct 2026; each link shows the current state.

ProjectChangeStatus
NVIDIA garakFixed a crash in z-rating evaluation caused by a removed calibration method. PR #1794 ↗Merged
NVIDIA garakAdded a regression test to keep duplicate configuration files out of scan reports. PR #1791 ↗Merged
PromptfooFixed evaluation configs silently dropping distinct function-based providers. PR #9402 ↗Merged
PresidioFixed a YAML default that scored deny-list matches at zero, silently missing PII at positive thresholds. PR #2254 ↗Merged
Anthropic Claude CodeCredential Guard: a proposed guardrail to detect hardcoded secrets in coding-agent tool calls. PR #62099 ↗Open PR
Anthropic Claude CookbooksProposed secure Claude Code routines for avoiding credential leakage. PR #667 ↗Open PR
Future AGIProposed an indirect prompt-injection guardrail for the AI gateway. PR #851 ↗Open PR
NVIDIA NeMo GuardrailsProposed synchronization for embedding batches. PR #1921 ↗Open PR
ProtectAI ModelScanProposed a fix for settings-file generation producing an empty file. PR #349 ↗Open PR
ETH AgentDojoProposed removing an unused processed-message list from the prompt-injection detector. PR #195 ↗Open PR
NVIDIA garakProposed structured CLI exit codes for automation and failure handling. PR #1790 ↗Open PR
NVIDIA garakProposed lint configuration and core-module fixes. PR #1793 ↗Open PR

Independent use

The Prompt Injection CTF was adopted by an engineer at Apiiro as the fixtures for a published nine-model study. The original Jack & Jill vulnerability was confirmed by its founder.