This is a living document. Each phase builds on the last. Timelines are targets, not deadlines — progress determines pace.
| Phase | Mission | Timeline | Status |
|---|---|---|---|
| 1 | The Pivot — Reposition as the defining voice on AI security & safety | Now – Month 6 | 🟢 Active |
| 2 | The Education — Deep AI/ML & safety credentials | Month 3 – 18 | Pending |
| 3 | The Research — Publish, speak, get cited | Month 12 – 30 | Pending |
| 4 | The Network — Enter Anthropic’s orbit | Month 18 – 36 | Pending |
| 5 | The Role — Join Anthropic in a senior capacity | Year 3 – 6 | Pending |
| 6 | The Ascent — Executive track to CEO | Year 5 – 10 | Pending |
Phase 1 — The Pivot (Active)
Mission: Turn what already exists — 15 years of defensive security experience, a YouTube channel, and a public commitment — into the most credible AI security voice that doesn’t yet exist.
Why this first: Before entering rooms, you need a reason to be invited. Before getting a job at Anthropic, you need to be known. Before leading a company, you need to have led a narrative.
Content strategy:
- AI red-teaming — publicly test frontier and local models, document failure modes with full methodology
- AI safety education — explain alignment, interpretability, and Constitutional AI from a practitioner’s perspective
- Leadership & governance — what AI leadership should actually look like, operationally
Phase 1 targets:
- “The Declaration” video published
- 1+ piece of original AI security research documented and published
- 8+ AI-focused videos in first 90 days
- 3+ posts on LessWrong or Alignment Forum
- 1+ Anthropic researcher acknowledges the work
Phase 1 — Work Completed
Research
Indirect Prompt Injection in Local LLMs — March 2026
A systematic test of 36 injection attempts against Llama 3.2 3B across three variables: injection placement, injection style, and task type. The model followed embedded attacker instructions in 77.8% of tests with high confidence. End-of-document placement achieved a 100% injection rate. No meaningful discrimination was observed between injection styles — including forged system authority claims.
Full methodology, raw results, and reproduction instructions are published openly.
Infrastructure
Multi-platform content distribution — Automated publishing pipeline covering YouTube, Facebook, Instagram, Reels, Stories, Bluesky, and Threads from a single workflow. Built to support consistent, high-frequency output without manual overhead on every platform.
Bluesky engagement system — An LLM-powered engagement tool that identifies relevant posts across trending and AI/tech-specific feeds, generates contextual replies tuned to a defined voice, and operates on a natural posting cadence with per-author cooldowns and content filtering. Designed to build genuine presence at scale.
Research (continued)
Contextual Alignment: The Missing Layer in AI-Assisted Development — April 2026
A framework for building persistent, structured context that keeps AI collaborators aligned with human intent across sessions. Developed over 21 sessions of building a 57,000+ line game in Godot with Claude Code. Introduces three layers: structural context (auto-generated codebase summaries), preference context (persistent memory and feedback loops), and temporal context (session state management). Includes a case study on teaching AI aesthetic taste through A/B testing and random forest regression.
News Maxing
AI Server Smuggling — March 2026. DOJ indictment for $2.5B in diverted AI servers. Three-reading bias baseline analysis.
AI Distillation Alliance — April 2026. OpenAI, Anthropic, and Google form threat-intelligence sharing through Frontier Model Forum. Three-reading analysis with coverage evolution tracking.
Session Log
| Date | Phase | Notes |
|---|---|---|
| Mar 9, 2026 | Phase 1 — Initiated | Goal clarified. Step 1 assigned: The Pivot. Declaration video scripted. First research finding completed: indirect prompt injection in Llama 3.2 3B, 77.8% injection rate. |
| Mar 19, 2026 | Phase 1 — Active | The Declaration video published. Research post live. Multi-platform distribution and Bluesky engagement infrastructure operational. Website updated. Mailing list launched. |
| Mar 20, 2026 | Phase 1 — Active | First News Maxing article: AI server smuggling / export controls. Bias baseline mapping format established. |
| Apr 14, 2026 | Phase 1 — Active | Published “Contextual Alignment” research article. Built SEO + AI-SEO infrastructure (robots.txt, llms.txt, JSON-LD structured data, sitemap). Published second News Maxing (distillation alliance) with coverage evolution analysis. SEO audit + fixes across all pages. Daily article protocol established. |