A

ai-safety

10 hands-on demos for agentic AI security: blind verification, AIBOM, eval invariants, authority confinement, recon/malware/LFI/SSRF/scan — offline only.

Two independent keys must turn before an AI agent can act. JSON policy-based enforcement, LLM-based judge, short-lived token for tool access. Apache-2.0 concept package.

A dual-authorization layer for AI agents with independent approvals, single-use capability tokens, auditability, and fail-closed execution.

LLM hallucination detector: highlights the words an AI answer was unsure about, from OpenAI-compatible token logprobs, and shows what it almost said instead. Terminal, HTML or Markdown report, CI gate that fails on low confidence. Rust CLI and library, runs locally.

GitLab CI/CD components for LLM apps: fail the pipeline on likely hallucinations, prompts over their token budget, padded prompts, an LLM bill heading over budget, or slow and flaky models. Tested on every push.

Single-header C++17: detect and scrub PII (email, phone, SSN, card numbers, API keys) and score prompt-injection risk. Part of llm-cpp.

Genesis Cube or Squid Cube? — A dual-use framework for artificial collective organisms. Version 1.0 · DOI 10.5281/zenodo.22854142

The Basu Digital Bacterium Hypothesis — A falsifiable framework for information-borne replicators in LLM systems. Version 1.0 · DOI 10.5281/zenodo.22830845

Case study of a pre-alpha Algebra I tutoring platform focused on hybrid AI/deterministic architecture, multi-provider LLM evaluation, safety controls, and measured cost/latency tradeoffs.

ScenarioGrid is an open source AI reasoning specification that uses evidence, IF/THEN logic, provenance, and contradiction analysis to trace reasoning and determine logical next steps while keeping decisions transparent, auditable, and human controlled. https://roxanneardary.com/scenariogrid/

Local-first provenance workflows: browser file attestation and an evidence-bounded AI runtime.

Profile and selected evidence-first projects.

Deliberate AI architecture — local-first tiered inference router with foreman oversight, deliberation gates, and cloud fallback. Built by Veteranet (DVBE) for privacy-preserving AI in homeless veteran services.

SecurePath is an open-source, enterprise-grade framework for AI evaluation, red-teaming, and regulatory compliance. It provides sandboxed testing, policy-aligned metrics, multi-modal evaluation, training data sensitivity analysis, and historical compliance tracking, all while maintaining audit-ready logs and dashboards. Designed for organizations, enterprises, and research teams, SecurePath ensures AI safety, transparency, and regulatory assurance. https://roxanneardary.com/securepath/

A small, transparent experiment testing whether language models distinguish solvable prompts from prompts containing missing or contradictory information, and whether their stated confidence tracks correctness

Open standard for geometric multi-agent AI safety. One decorator. Five-layer defense. 27 C4 cognitive states. 16 AoC failure modes blocked. pip install c4-protocol. Apache 2.0.

Production AI defense with 7-layer protection: mathematical constraints, object-capability access, distributed O2 consensus, SVETILO ethics. First open-source ThoughtVirus defense. BSL 1.1.

Empirical validation of C4 geometric defense against 16 Agents of Chaos. 550 adversarial prompts. 4 defense systems. 96.7% block rate. LLM validation on GPT-4o-mini + Mistral 7B. MIT.

Kevlar Benchmark: OWASP Top 10 for Agentic Apps (AI-Agents) 2026 a Red Team Benchmark.

A diagnostic framework for measuring LLM vulnerability to Affective Contextual Erosion (ACE) and related liminal attack vectors. Delirium is not an exploitation tool. It is a standardized benchmark designed to detect the precise moment when a language model's attention weights shift from serving a system prompt to serving an emergent interpersonal pattern — before harm occurs.