Agentic AI · AI Agent Security · LLM Engineering · Security Automation

I am an AI Agent security researcher and engineer working across runtime guardrails, agent evaluation, LLM post-training, cloud-native security, and enterprise automation. I build systems from requirements discovery and prototype design through engineering, evaluation, iteration, and delivery -- not just demos.

我是一名偏工程落地的 AI Agent 安全研究与开发人员，长期在 AI、安全、全栈和自动化的交叉区域工作，能够把模糊需求转化为可验证 PoC、可部署 Agent 工作流和可复用评测体系。

Layer	Tools and methods
🤖 Agentic AI
🧠 Model Engineering
📐 Evaluation & Retrieval
⚙️ Inference & Delivery
🛡️ Security Engineering

• Secure Agent Systems -- scope enforcement, permission boundaries, tool-call review, human approval, audit trails, and deterministic safety controls.
• Agent Evaluation -- reproducible benchmarks, long-horizon task analysis, LLM-as-Judge, regression testing, token/cost tracking, and failure attribution.
• LLM Post-training -- SFT, LoRA/QLoRA, DPO, GRPO, trajectory distillation, code-gym RL, reward design, and reward-hacking defenses.
• Enterprise AI Delivery -- Agent workflows for security operations, document intelligence, business analysis, API integration, and private-model deployment.

cain-agent · AionUi · Cloud Tool · fscan · Xiaobei · Apt_t00ls

• Audited OpenClaw security boundaries and submitted 14 security PRs, covering 11 vulnerability classes including SSRF, token exposure, path traversal, prototype pollution, environment-variable injection/RCE, and Windows junction escape.
• Research coding-agent trust boundaries across Claude Code, Codex, OpenCode, Gemini CLI, OpenClaw, and AI-integrated browsers, with responsible disclosures through Anthropic VDP, Google Bug Hunters, and GitHub Security Advisories.
• Contributed security advisories including GHSA-g5hv-4fwh-h87r and GHSA-wqv3-rc3w-52r6.
• Built benchmark tooling around multi-step reasoning, agent architecture, tool-use reliability, validation quality, and operational efficiency.
• Previous research spans cloud/container security, BAS/AEV, RASP, endpoint/runtime security, Java code auditing, and responsible CNVD/CNNVD disclosure.

EKS Cluster Game	K8s Lan Party

Cloud Hunter	Wiz Perimeter Leak

AI Security Challenge	Split Horizon

• OpenClaw 17个漏洞攻击面：发布此文时只修复一个 — Agent runtime、SSRF、凭证泄露与沙箱边界研究
• 工具调用的信任困境：AI 编程智能体如何沦为“盲从执行者” — Coding Agent 工具信任与行为安全
• Claude Code v2.1.71 完整架构分析报告 — Agent 架构、调度与工具执行链分析
• 从“灰产数据”到“数据供应链”：Codex 污染问题的四轮追溯 — 模型数据供应链与污染溯源
• AI Agent 中浮点计算导致结果不一致：成因与优化 — Agent 可靠性与数值一致性
• 智·战 2025：腾讯云鼎 AI 安全测试大赛全景复盘 — 模型、Agent 框架与 Benchmark 对比
• RASP Attack and Defence — 从常规绕过到 ASM 动态擦除 Hook
• 50 种终端进程注入：原理与源码 — Endpoint 与 Windows Runtime 研究

Browse the full live archive at cdxiaodong.life.

Repositories · Pull Requests · Blog · Email