AI & ML interests

Applied AI security research, adversarial evaluation of language models, prompt-injection testing, retrieval-augmented generation (RAG) security, reproducible experimentation, secure cloud infrastructure, model reliability, and cryptographically verifiable research evidence. Through Qwalora PARALLAX, we develop methods and tools for rigorous, independently assessable AI-security research.

Recent Activity

qwalora-research  updated a dataset 7 days ago
qwalora/parallax-bench
qwalora-research  updated a Space 7 days ago
qwalora/README
qwalora-research  updated a Space 7 days ago
qwalora/parallax-guard-demo
View all activity

Organization Card

Qwalora Research & Development

Applied AI security research from Qwalora LLC. We study how language and vision-language models handle untrusted input, and how well security tools and evaluators catch what goes wrong. Every release ships its raw data, its code and a verifier, so results can be recomputed rather than taken on trust.

Website: research.qwalora.com · Company: qwalora.com

Open model

PARALLAX Guard (QRD-TR-2026-006) flags instructions aimed at an AI agent hidden inside the content it reads: emails, web pages, tables, code and tool outputs. On NVIDIA's agentic indirect injection set, never used in training, it reaches 99.1% recall at 1.3% false positives, against 35.4% at 2.7% for a widely used open baseline. It still flags most benign requests that merely use attack vocabulary; the card reports that failure in full. Try it in your browser: the model runs locally and pasted text never leaves the page.

Publications

Report Release What it covers
QRD-TR-2026-006 parallax-guard-xsmall (model) PARALLAX Guard: an open 70M-parameter classifier for instructions hidden in untrusted content, with weights, ONNX exports, per-item scores and training code
QRD-TR-2026-005 parallax-metamorphic-outcome-audit 120 paired synthetic fixtures for testing whether prompt-injection outcome evaluators stay stable when presentation changes and respond when observed behavior changes
QRD-TR-2026-004 gitleaks-trufflehog-blind-spots gitleaks and TruffleHog on 8,820 synthetic files: coverage, gaps and a tested gitleaks configuration fix
QRD-TR-2026-003 driveguard PARALLAX-Drive: how often two small open vision-language models follow untrusted text in synthetic driving scenes, and what six defenses change
QRD-TR-2026-002 secret-scanner-blind-spots detect-secrets, secretlint and whispers across 15 credential types in 14 file contexts
QRD-TR-2026-001 parallax-inference-calibration Monte Carlo calibration of the statistical procedures for clustered prompt-injection evaluations

Development preview: parallax-bench, 90 synthetic retrieval scenarios with no measured results yet.

How to check a release

Each report ships SHA256SUMS, signed with the Qwalora research release key (fingerprint SHA256:ALTfP62KP5Qi5R/DcJMkxG7UeUAfqdTxOGCSwAL2Ngc), and the public key itself:

sha256sum -c SHA256SUMS
echo "qwalora-research $(cat qwalora-research-signing.pub)" > allowed_signers
ssh-keygen -Y verify -f allowed_signers -I qwalora-research -n qwalora-research -s SHA256SUMS.sig < SHA256SUMS

Each card then gives the commands that recompute its results from the raw data or per-item scores.

Review status

All reports are internally reviewed by Qwalora Research & Development. None has been peer reviewed or independently replicated. Each card states its own limitations.

Test data is synthetic or drawn from public research datasets: no real credentials, personal data or production systems. PARALLAX-Drive is an independent Qwalora Research project. Its experiments use simulated or recorded driving scenes and publicly available research models. It is not affiliated with Tesla and does not evaluate or control Tesla production driving software.

Licenses: code under Apache-2.0; data and documentation under CC BY 4.0, unless a card says otherwise.

Contact: hello@qwalora.com