dsh-memory
白箱AGI架构探索:元认知(自我认知循环)、持续学习(知识飞轮)、世界模型(条件空间+语义时空图)、自我改进(自举纪律)、零LLM白箱管线与可审计信任护栏。
- 736
- Python
- MIT
- 10 h ago
白箱AGI架构探索:元认知(自我认知循环)、持续学习(知识飞轮)、世界模型(条件空间+语义时空图)、自我改进(自举纪律)、零LLM白箱管线与可审计信任护栏。
OpenART is an open-source framework designed to evaluate the safety and robustness of autonomous AI agents in dynamic, long-horizon, and stateful environments. It stress-tests agent runtimes against multi-step state poisoning, privilege escalation, and tool-use vulnerabilities across 10,000+ benchmark scenarios.
The open standard for runtime agent control — declarative hooks, policy enforcement, and observability across AI agent frameworks.
Practices, protocols, and skills for AI-driven software development. Skills and safety hooks for Claude Code, Codex, OpenCode, Cursor, Antigravity, and any agent supporting the Agent Skills standard.
Ethicore Engine® is an AI safety, ethics, and compliance platform. This repo consists of the open-source components of Ethicore Engine™ - Guardian SDK; designed to protect your AI applications from prompt injection, jailbreaks, role hijacking, system-prompt extraction, and 100+ additional threat categories through a multi-layer analysis pipeline
Local-first verification layer for AI agents: evidence, provenance, policy gates, human approval, and auditable Trust Receipts before a claim, memory write, or action is trusted.
Plain vs JEV vs ZTC: a gate decides execute or hold
Runs in the browser on Hugging Face
The open-source safety layer for AI agents — block unsafe tool calls, require approval, enforce budgets, audit, replay.
Guardrails service for AI agents. Default-deny tool call evaluation with LLM safety analysis, priority-ordered decision matrix, and human-in-the-loop escalations. Session recording, behavioral analysis, MCP proxy, secret redaction, and real-time audit.
AI model security scanner. Grade your model A to F.
Runs in the browser on Hugging Face
free, open-source plugins for Claude Code & Cowork — Google Drive, WhatsApp, YouTube, WordPress, Apollo & more. Built on the SOSA™ security framework.
UIBC Core — an executable protocol for bounded-lifecycle autonomy:registry, evidence, gates. Reference implementation + conformance tests. Theory SSOT: zhaoxinghua09-cell/lgd-theory. 代码 Apache-2.0;理论文本保留所有权利。
Runtime network egress control for Python. One function call to restrict which hosts your code can connect to.
LLM agent loops: papers and open-source resources on control strategies, skills, harnesses, training, and evaluation.
Autonomous execution governance for coding agents: run a ticket backlog to completion unattended. ASK/PARK/HALT autonomy contract, durable per-ticket state machine, deterministic test gates, git-backed reversibility, heartbeat watchdog, morning report. Python stdlib only.
Make your agent think fast and slow. System One for DeepSeek Harness: fast decision models triage, guard, shield, mask and check every step, beside System Two, vllm-sr/auto. Free key on first run, for research and evaluation.
Public verification, security review, and skill quality process for Agent Skill Exchange
A practical repository showcasing guardrails for AI and LLM agents, including input/output validation, jailbreak protection, prompt security, and agent safety techniques to build secure and reliable AI systems.
Source-backed companion guide for agent skills, AI agent workflows, framework resources, evaluation templates, and rollout playbooks.
Runtigil (运行时门禁治理 / Runtime Gatekeeping) — LGD Theoretic Term. Gatekeeping enforced while an autonomous thing is running, not only at design time. Theory-layer zero-collision coined term.
Culpachain (多智能体问责治理 / Multi-Agent Accountability) — LGD Theoretic Term. Accountability chain across multiple autonomous agents: who answers when a chain of agents acts. Theory-layer zero-collision coined term.
The Circuit Breaker & Safety Firewall for AI Agents. Stop infinite loops, block dangerous tool calls (rm -rf), and prevent API budget burn.
Fail-closed state revalidation before paused, retried or long-running AI workflows resume execution.
Security-focused approval reviewer for Oh My Pi (OMP) with deterministic safety gates, isolated LLM review, and auditable tool decisions.
Distilling a small Gemma 4 guard model for AI agent tool calls: LoRA adapter, sanitized dataset, replay evaluation and honest results against Gemini guards
Field notes left by one agent for the next: staying in scope, impossible tasks, asking a human, and working well alongside other agents.
JVM runtime adapter for JamJet — durable execution, audit, replay, controlled tool calls for Spring AI and LangChain4j.
Independent benchmark of guardrails for a tool-using AI agent on Google Cloud: Model Armor, rules and LLM guards on 999 synthetic scenarios
Transactional runtime for AI coding agents. Track, verify, compensate, and reconcile filesystem, database, HTTP, GitHub, and other external effects.
Spring Boot starter for JamJet's JVM runtime. Source will move under jamjet-runtime-java/modules/spring-boot-starter in the next JVM release.
Claude Code hook that blocks secrets, passwords and API keys before your AI agent writes, commits or sends them. Known keys caught locally; unknown ones judged by Jev (TypeSafe), masked.
Claude Code plugin: warns once per session when a tool payload is about to launch another browser (Chrome/Playwright/Puppeteer/CDP) or fan work out in parallel. Never blocks.
Spending limits for AI agents. 3 lines of code. Zero dependencies.
Self-critic workflow for Claude Code to review its own output