I build production-grade LLM agent systems with LangGraph, run multi-agent reinforcement learning (MARL) for cybersecurity defense, and operate adversarial red-team tests against AI systems. Currently @ DBAPPSecurity.
- 🌐 Blog: appmiaoda.com
- 🏢 Company: dbappsecurity.com
- 🧪 Orgs: MARL-ECDSA-Consensus · Circle1507
Active contributor to the Octop agent platform (CLI · harness · memory). I'd rather point at merged work than quote counts — the table below lists changes that were reviewed and merged by maintainers, with links you can click and check.
Profile forks (synced or ahead of upstream):
- TrueFurina/Octop → upstream main (synced)
- TrueFurina/octop-memory → upstream main
- TrueFurina/octop-harness → 2 commits ahead of upstream
只列已被上游合并的改动，每条都能点开核对（维护者原话在 PR 里）。
| 上游 / PR | 改了什么 | 规模 | 结果 |
|---|---|---|---|
| unicode-rs/unicode-segmentation #183 | GraphemeCursor 公共 API 的越界 panic：offset/chunk_start 切片、prev_chunk 反向迭代下溢、pre-context 分块校验 |
+27/−5 | 维护者 Manishearth 要求 cargo fmt 后 APPROVED 并致谢 |
| CodeGraphContext #1676 | 给 Cassandra / MySQL ingestor 加 SSL 连接参数（CERT_REQUIRED + 主机名校验，默认关闭、不影响旧配置） |
+1214/−315，25 文件 | 维护者本地 trial-merge 验证后合并 |
| TencentCloud/Octop #2171 → #2174 | 报 issue：OCTOP_HOME 不迁移向导密码文件；逐行指出 4 处该同步修改的集成测试，维护者据此修复 |
+173/−40 | 维护者 "Looks merge-ready for #2171"，changelog 署名 (#2171) |
| seanmonstar/reqwest #3100 | base64 改 default-features = false，去掉 simd-unsafe，crate 恢复 #![forbid(unsafe_code)] |
+1/−1 | 作者 seanmonstar 本人 APPROVED |
| Unstructured-IO/unstructured #4463 | Text 元素在 None 文本时归一化为空串，修 TypeError（含测试） |
+33/−4，4 文件 | 维护者 APPROVED |
我更擅长定位而不是堆代码：Octop 那条我没有直接改代码，而是把维护者需要同步修改的 4 处测试行号列清楚，他按清单改完就合了。
| ⭐ | Repo | Domain |
|---|---|---|
| 8 | AGI-Distiller | 🧠 Knowledge distillation |
| 6 | passive-recon | 🛰️ OSINT / EASM |
| 5 | MARS-408 | 🎓 MARL |
| 4 | xihu-sword-ctf-agent | ⚔️ CTF AI agent |
| 0 | marl-ecdsa | 🔐 ECDSA consensus via RL |
| 3 | xuandun-cyber-range | 🎯 AI RedTeam range |
LLM Agents
├─ LangGraph pipelines (plan → act → observe → reflect)
├─ RAG optimization (retrieval quality vs latency)
└─ Knowledge distillation (small ≈ big?)
Multi-Agent RL
├─ Cooperative consensus (ECDSA, PBFT via RL)
├─ Adversarial games (cyber attack/defense as zero-sum)
└─ MARL × LLM hybrid (LLM as policy network)
AI Security
├─ Red-team automation (jailbreak benchmarking)
├─ Prompt injection defense
└─ AI agent supply chain security
- 🤖 SecAutoMind — LLM-powered automated security testing pipeline
- 🧪 MARL for consensus — ECDSA signing without a leader
- 🛡️ AI red-team framework — systematic adversarial evaluation
- 🧠 Agent memory layer — cross-session persistence
| Layer | Tools |
|---|---|
| Agent Frameworks | LangGraph · LangChain · AutoGen · Dify |
| LLMs | DeepSeek-V3 · GPT-4o · Claude-Sonnet · Qwen |
| MARL | PettingZoo · Ray RLlib |
| Security | ZAP · Nuclei · SpiderFoot · Promptfoo |
| Backend | FastAPI · Docker · PostgreSQL |
Code first, buzzwords never.