AI Security Researcher · Open Source Contributor · Taiwan
I build tools that protect AI systems from prompt injection attacks. 15 PRs merged across seven organizations: Microsoft, NVIDIA, Cisco, NIST, OWASP, the UK AI Security Institute, and the Casualty Actuarial Society.
Before this I spent ten years as a financial adviser (MDRT), which is why the actuarial and finance rows below exist: I read insurance code as a practitioner, not a tourist. I run Ultra Lab, an AI products company in Taiwan.
My niche is unglamorous: I develop on Windows with a Traditional Chinese (cp950) locale, an environment almost no OSS contributor uses. Bugs that are invisible on Linux CI live there for years. Most of the merges below started that way.
Every row links to the merged work; verify any of it with gh pr list --author ppcvote --state merged.
| Organization | Merged | What |
|---|---|---|
| Microsoft · PyRIT / agent-governance-toolkit | 5 | OWASP LLM02 output-scorer pack; PromptDefenseEvaluator |
| OWASP · regression harness / AI testing guide | 3 | Agent-security scenarios; testing-guide mapping |
| NVIDIA · SkillSpector | 2 | Windows silent file-skip race; LP3 remediation guidance |
| UK AI Security Institute · inspect_evals / inspect_k8s_sandbox | 2 | Eval config migration; K8sError diagnostics |
| Casualty Actuarial Society · chainladder-python | 1 | Quarterly/semiannual premium on-leveling crash |
| Cisco AI Defense · mcp-scanner | 1 | PromptDefenseAnalyzer prompt hardening |
| NIST · dioptra | 1 | Task-plugin path convention docs |
In review: NASA (forced-colors focus ring) · MITRE (UTF-8 changelog) · CERT/CC (approved) · FINOS (gitleaks range corruption on Windows) · De Nederlandsche Bank (input-frame mutation) · OpenSSF (17/18 language checkers dead on Windows)
Pre-deploy scanner — checks system prompts for missing defenses. 17 attack vectors, SARIF 2.1.0 output. npx prompt-defense-audit "Your prompt"  | Runtime scanner — blocks injection attacks before they reach your LLM. 8 attack types. npm i @ppcvote/prompt-shield  | CI/CD gate — auto-scans prompt files on every PR. Posts results table. - uses: ppcvote/prompt-defense-audit-action@v1 |
Pre-deploy → CI/CD → Runtime
prompt-defense-audit GitHub Action prompt-shield
"Does your prompt have "Block weak prompts "Is this user input
defenses?" before merge" an attack?"
Scanned 1,646 production system prompts from ChatGPT, Claude, Grok, Copilot, and 1,300+ GPT Store apps:
| Finding | Rate |
|---|---|
| No indirect injection defense | 97.8% |
| No role boundary enforcement | 92.4% |
| No output control | 88.3% |
| Overall grade F (score < 45) | 78.3% |
| Average defense score | 36/100 |
Data: gap-20260405.json · Paper: DOI 10.5281/zenodo.19410475