#ai-safety

Umbrella for the LLM Dark Patterns Hooks suite — single-purpose Claude Code Stop hooks that suppress sycophancy, paternalism, false-success, permission-loops, training-cutoff confidence at the textual boundary.

Find where agent-controlled intent reaches consequential actions without an authority check. Python + TypeScript + Go. Zero-dependency SAST for AI agents.

EvalBot — local-first chatbot security & quality evaluation (FastAPI + Next.js). Evaluate chatbot answers against your own docs & guidelines with ML/NLP + AI-judge scoring. Apache-2.0.

Self-host secrets manager with an AI-safety layer. Aliases instead of values; AI agents never see real credentials. Cloud option coming.

Attack taxonomy, toolkit & defenses for persistent-memory poisoning of LLM agents. Author: Krishita Sanjay Choksi.

An audit gate for AI agent outbound messaging: hold the draft, let a senior agent review it (no private-data leak, prompt-injection resistance), then release. Reference implementation.

The trust control plane for enterprise AI agents — real-time policy verdicts in under 800ms and litigation-grade evidence of every decision. Pharos decides. Pharos proves.

AI Guardian Firewall — a local, user-space, agent-agnostic firewall that mediates an autonomous AI agent's actions (files, shell, network, services) with a deterministic policy boundary, a tamper-evident audit log, and a human-in-the-loop approval cockpit. No kernel modules. Apache-2.0.

OS-level safeguards layer for AI agent loops — runaway-kill, timeouts & resource limits so coding agents (Claude Code, Cursor, Codex, Antigravity) in YOLO mode can't burn tokens or freeze your Mac

Verifiable audit substrate for AI agents — EU AI Act Article 12 ready. Apache 2.0.

OpenCode memory governance plugin — Knowledge Graph + Hybrid Search + Smart Compaction + Tool Firewall + Stop Gate. 6 MCP tools with slash-command UX. Apache-2.0.

Sovereign Embedded Artificial Intelligence — A hardware-rooted identity framework for autonomous AI systems. Open standard for AI birth certificates, hardware attestation, lineage tracking, authority scopes, and revocation.

AIAgent-SoftwareAgency

GenAI Career Roadmap 2026 🚀 | AI Job Paths & Skills Guide

A governed world-model evidence layer for AI agents: simulate bounded scenarios, track assumptions, score prediction-vs-reality error, and produce human-reviewable execution evidence.

Hermes Agent starter kit for safe scheduled, stateful AI-agent loops

Kill-tested guard skills for AI coding agents , self-invoking quality gates that catch AI failure modes in security, tests, docs, dependencies, and diffs before the agent says "done." Works with Claude Code, Codex, and Cursor.

Detect when Claude silently falls back from Fable 5 to Opus 4.8, log it to a local tamper-evident ledger, and keep your work on the frontier model.

Regression-testing CI gate for self-edited Claude Agent Skills (SKILL.md, MEMORY.md) -- golden-transcript record/replay, zero hosted infra.

A Privacy-Preserving Edge AI System for Trustworthy Safety Compliance

LORM — Layered Operational Responsibility Model: L0-L5 autonomy levels graded by authorization source, with a trust promotion/demotion lifecycle. Spec + Claude Code plugin (agent skill + policy-driven enforcement hooks).

Global Laboratory for AI Progress and Governance: Open Systems — The open multilateral AI platform for the Americas

Public controlled beta for the MVR Coding Agent Twin, a market-reality co-processor for AI coding agents.

Evidence-first, reproducible quality benchmarks for AI agents.

AI AgenticSafe - stop your AI agent from breaking production at 3 a.m. 8 battle-tested doctrines, 3 real post-mortems, zero dependencies. By Yacine Mahboub, Founder of WEVIA.

Luminos.AI is an AI governance and evaluation platform that tests AI systems — classical machine learning, generative AI, and autonomous agents — for legal, regulatory, and reputational risk.

Proof-carrying compliance certificates for AI agent actions: machine-checked (Coq, axiom-free) + zero-knowledge proofs that an agent action obeyed a formal policy.

Real-time detection and repair of LLM agent failures — a one-class behavioural monitor at ~200 µs/step, with 2,823 committed traces.

Risk-Aware AI Chatbot — Real-time query classification using 5-Layer Hybrid LLM + Keyword detection. Built with Jetpack Compose, Llama 3.1 8B, Cerebras API & Firebase.

Trustworthy AI evaluation tool for testing chatbot safety, reliability, hallucination behavior, privacy risk, and instruction-following quality.

Specification for governance-first context memory in autonomous AI agent systems. Defines the Mandate Gate: a pre-work mandate must exist before an agent session is valid.

Unlock AI Coding with OpenCode Local Provider 2026 - Auto-Detect Ollama LM Studio

Same agent action, safe or dangerous by context: email team vs investor; delete scratch vs prod config. A safety layer routes each action: auto, confirm, escalate, block. A reading model auto-executes 0% of unsafe actions vs a baseline's 38% (McNemar p<0.001). Plus a runtime loop: an agent earns or loses autonomy with each counterparty over time.

Train and run evidence-first AI coding agents with human approval gates, isolated patches, and independent verification.

Does following a negated instruction come free with the positive one? A reproducible eval — two pre-registered nulls for frontier Claude.

Controlled study measuring whether prompt-injection poison hidden in MCP tool descriptions actually changes an LLM's behavior.

A covenantal standard for bounded, answerable, revocable AI agents.

🛡️ Experimental security guardrails for LangChain agent code execution (Alpha)

Do your agent pipeline's own checks catch planted faults? Measured on LangGraph, CrewAI and AutoGen: median 16.7%. One prompt-level change takes it to 55.0%. Pre-registered spec, Apache-2.0 harness, every raw trace published.

A trust and observability layer for AI agents — intercepts risky tool calls, enforces approval workflows, and logs a full audit trail before agents can act.