|
Pre-tool-call governance + structured audit for OSS agent frameworks. Adversarial security review of every TypeScript package I ship. A mechanism-grounded jailbreak taxonomy that publicly retracts overclaims instead of silently rewriting them. |
Custom bot builds (Discord / Telegram / Slack — AI-powered or classic) · web builds (Next.js, Astro, Three.js / R3F) · community moderator + support-agent roles · AI-safety consulting · independent security review for TypeScript + Node 22 · Three.js + WebGPU performance audits · technical writing with sourced citations. |
Verification-first TypeScript agent framework. Hallucination defense, source grounding, pre-tool-call governance, capability-token sandboxing, and platform rate-limit scheduling as first-class layers — not bolt-ons.
|
Mechanism-grounded taxonomy of 40 LLM jailbreak patterns across 10 categories, mapped to the safety-alignment assumptions they subvert.
|
The Three.js + WebGL performance reference I wish I'd had: 48 validated topic folders, every claim sourced against live repos and browser specs.
|
AI / Safety · Anthropic SDK · OpenAI SDK · Constitutional AI · RLHF · red-teaming · Wilson CI · McNemar · Cochran Q
Agents · TypeScript strict · ES2024 · pnpm workspaces · MCP · ACP · OAuth gateway · wasmtime sandbox · multi-judge verifier · pre-tool-call governance + audit
3D / Web · Three.js · React Three Fiber · drei · WebGPU · WGSL · GLSL · GSAP · Lenis · Vite · Astro · Next.js 15 · Core Web Vitals
| Date | Project | Summary |
|---|---|---|
| 2026-06-04 | TENET |
Phase 14 vulnerability-test triage — three-agent adversarial review surfaced 23 real issues; HIGH + MEDIUM patched (a2ca9a8) |
| 2026-06-04 | @tenet/acp |
Agent Client Protocol v1 adapter — JSON-RPC 2.0 over stdio, NDJSON framed, for Zed / Cursor / Helix interop |
| 2026-06-02 | llm-jailbreak-taxonomy |
v4.2.1 released with public retractions of four overclaims |
| 2026-06-02 | tldr-pages/tldr |
pyinstaller merged upstream · fc-scan / syft / helmfile PRs approved |
External-PR throughput this year: 12 PRs · ~850 commits
Sources: primary documents first (official docs, browser specs, live repos); peer-reviewed papers second. Affiliate review sites, paraphrased rules, and training-data memory are rejected outright.
Claims: hypothesis until corroborated by ≥2 independent sources. A single source is marked
UNVERIFIED. Retractions are documented publicly in the CHANGELOG — never silently rewritten.Simulation vs measurement: hand-tuned parameters that reproduce literature distributions are a prior, never presented as measurement. Live API calls against real models are the only valid measurement.
Review: adversarial review is welcome. The assumption is the reviewer is right until proven otherwise. Response order: fix → defend → silently ignore (last is never acceptable).
