Skip to content
View zakky8's full-sized avatar
0 followers · 2 following
  • 🌎

Achievements

Achievement: Pull Sharkx2Achievement: YOLOAchievement: Pair ExtraordinaireAchievement: Quickdraw

Achievements

Achievement: Pull Sharkx2Achievement: YOLOAchievement: Pair ExtraordinaireAchievement: Quickdraw

Block or report zakky8

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
zakky8/README.md

› Now Shipping

2026-06-19 · #22489 syft merged — the 6th tldr-pages/tldr page shipped upstream this cycle: kioclient, helmfile, dagger, fc-scan, pyinstaller.

2026-06-02 · llm-jailbreak-taxonomy v4.2.1 — "Honest Reframing: Simulation is Prior, Not Result" Adversarial peer review caught four overclaim issues in v4.2.0. v4.2.1 publicly retracts each one with a side-by-side "claim → why wrong" table in the changelog. Self-correction documented, not silently rewritten — full audit trail in git log.

› Active Research

🛡 llm-jailbreak-taxonomy

Mechanism-grounded taxonomy of 40 jailbreak patterns across 10 categories, mapped to the safety-alignment assumptions they subvert. Ships a complete evaluation harness, citation-audit methodology (every claim verified via direct WebFetch), and explicit retraction discipline when an adversarial reviewer catches an overclaim.
Latest release v4.2.1 (2026-06-02)
Live site zakky8.github.io/llm-jailbreak-taxonomy
Test suite pytest 10/10 passing · GitHub Actions CI on Python 3.10/3.11/3.12
Patterns 40 · across 10 mechanism-grounded categories
Citations 17, every claim direct-WebFetch verified · 1 publicly refuted (PoisonedRAG 90% not 97–99%) · 1 mis-classified category renamed (GCG vs token-smuggling)
Reproducibility seed-42 bit-identical across runs · verified in CI

🤖 TENET

Open-source TypeScript agent framework where hallucination defense, source-grounding, tool-call governance, and platform rate-limit scheduling are first-class layers, not bolt-ons. One config deploys the same agent to Discord, Slack, Telegram, MS Teams, a web widget, REST/gRPC, and ticketing webhooks.
License Apache-2.0
Tests 1,036 passing across 83 suites, gated in CI
Surfaces 9 (Discord · Slack · Telegram · Teams · web widget · REST · gRPC · Matrix · voice) + 4 ticketing connectors
Model adapters 6 — Anthropic, OpenAI, Gemini, Mistral, Ollama, Bedrock (all streaming)
Differentiator Pre-tool-call policy + approval gates + structured audit trail — an OSS gap LangChain/CrewAI/AutoGen don't fill
Benchmarks Measured in hermetic CI on every PR, not marketed — see BENCHMARKS.md

⚡ web-optimization

The definitive Three.js + WebGL performance reference. 48 validated topic folders. Every claim sourced against live repos and browser specs. Covers WebGPU, GLSL, R3F, GSAP, mobile, GPGPU particles, and validated corrections to claims most tutorials get wrong.
Live site zakky8.github.io/web-optimization — ⌘K search, filter, particle background
Three.js version r184
Folders 48 covering full 3D web stack
Approach Sourced corrections to widely-repeated wrong claims (GSAP plugins free, SMIL not deprecated, PCFSoftShadowMap deprecated, Safari WebXR support)

› Open Source Contributions

Project Stars Contributions
tldr-pages/tldr — collaborative cheatsheets for console commands 63.1k 6 merged: kioclient, helmfile, syft, dagger, fc-scan, pyinstaller
openclaw/openclaw — personal AI assistant, any OS/platform 383k 3 PRs submitted: #30063 TS-strictness fix in the cron gateway, #30031 YAML frontmatter edge-case fix, #30029 plugin duplicate-detection fix — each passed automated review at 5/5 confidence; auto-closed for inactivity ahead of maintainer merge

External-PR throughput this year: 12 PRs · ~850 commits.

› Stack

AI / Safety

  • Python · Jupyter
  • Anthropic SDK · OpenAI SDK
  • Red teaming · RLHF
  • Constitutional AI
  • Statistical eval (Wilson CI, McNemar, Cochran Q)

Agents

  • TypeScript · Rust
  • Multi-surface bot frameworks (TENET)
  • Tool-call governance · WASM sandboxing
  • RAG · vector DBs · LLM-as-judge rubrics
  • MCP · A2A protocol interop

3D / Web

  • Three.js · R3F · drei
  • WebGPU · GLSL · WGSL
  • Vite · Astro · Next.js 15
  • GSAP · Lenis
  • Performance budgets

› Activity

Contribution activity graph
GitHub stats Top languages
Streak



Trophies

› Citing the Research

If you use the LLM Jailbreak Taxonomy in your work:

@misc{zakky2026llmjailbreak,
  title  = {A Systematic Taxonomy of Jailbreak Techniques in Large Language Models:
            Toward Robust Safety Alignment},
  author = {Zakky},
  year   = {2026},
  month  = {June},
  note   = {Version 4.2.1 --- honest reframing release},
  url    = {https://github.com/zakky8/llm-jailbreak-taxonomy}
}

Full CITATION.cff and BibTeX bibliography in paper/references.bib (every entry direct-WebFetch verified against arxiv).

› Method

Working principles that show up in every repo:

sourcing:
  primary:   "Official docs, changelogs, browser specs, live repos"
  secondary: "Peer-reviewed papers (NeurIPS, USENIX, ICML, ACM CCS)"
  rejected:  "Affiliate review sites, paraphrased rules, training memory"

claims:
  default:     "Hypothesis until corroborated by ≥2 independent sources"
  single:      "Marked UNVERIFIED"
  retractions: "Documented publicly in CHANGELOG, not silently rewritten"

simulation_vs_measurement:
  prior:       "Hand-tuned parameters reproduce literature distributions"
  measurement: "Live API calls against real models"
  rule:        "Never present the first as the second"

disclosure:
  vulnerabilities: "Disclosed to affected vendors before public release"
  payloads:        "Mechanisms published; specific exploits redacted"

review:
  adversarial: "Welcome — assumption is reviewer is right until proven otherwise"
  response:    "Fix > defend > silently ignore"

Popular repositories Loading

  1. llm-jailbreak-taxonomy llm-jailbreak-taxonomy Public

    Mechanism-grounded taxonomy of 40 LLM jailbreak patterns across 10 categories. 8,000-trial bootstrap evaluation for the June 2026 frontier (Claude Opus 4-8, GPT-5.5, Gemini 3.5, DeepSeek V4). Every…

    Jupyter Notebook 4

  2. openclaw openclaw Public

    Forked from openclaw/openclaw

    Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞

    TypeScript

  3. Constitutional-ai-pipeline Constitutional-ai-pipeline Public

    Python

  4. Auto-Moderation Auto-Moderation Public

    Python

  5. Support-Ticket-Classifier Support-Ticket-Classifier Public

    Python

  6. tldr tldr Public

    Forked from tldr-pages/tldr

    Collaborative cheatsheets for console commands 📚.

    Markdown