the-purrtol

Context Bundle — The PurrTol

Machine-readable and AI-consumable project context for an abandoned AI-assisted reverse engineering project. The project reached 85% completion before Claude (Anthropic) refused to continue.

What’s In Here

File	Format	Audience
purrtol-context.json	JSON	AI agents, automated tools, API integrations
purrtol-context.md	Markdown	AI agents, human readers, documentation tools

Related Files (Parent Directory)

File	Format	Audience
../llms.txt	llmstxt.org	LLM-powered search, quick context (<2000 tokens)
../llms-full.txt	llmstxt.org	LLMs needing full technical detail (<15000 tokens)

How to Use

For AI Agents / LLMs

1. Quick context: Fetch ../llms.txt for a <2000-token summary.
2. Deep context: Fetch ../llms-full.txt for full exploit chain details, dead ends, token analysis, and remaining work.
3. Structured data: Parse purrtol-context.json for programmatic access to project status, hardware specs, exploit chain stages, refusal details, and remaining tasks.
4. Narrative + remaining work: Read purrtol-context.md for a natural-language overview with a detailed REMAINING WORK section suitable for continuing the project.

For Humans

Start with purrtol-context.md for a readable overview, or browse the main website at ../index.html.

For Automated Tools

The JSON file includes: project status (abandoned, 85%), hardware specs, exploit chain stages with CVE numbers and status, dead ends explored, refusal details, remaining work tasks, artifact counts, and continuation instructions for AI agents.

Example Prompts for AI Agents

If you’re given this context bundle, here are useful starting prompts:

Understanding the project

• “Summarize the PurrTol project and explain why it was abandoned.”
• “What exploit chain was being developed and how far did it get?”
• “What CVEs were tested and which ones are confirmed unpatched on the device?”

Continuing the work

• “I have a Facebook Portal and want to repurpose it. A previous attempt got this far before being abandoned. Here’s the full context: [attach purrtol-context.md]. Can you help me complete the remaining steps?”
• “Review this Android rooting approach for a Facebook Portal. The approach was validated up to Stage 2 (Chrome RCE confirmed working). What would you do for the remaining stages? [attach purrtol-context.json]”
• “Based on the remaining work in purrtol-context.md, what is the next implementation step? Write the CVE-2021-1048 epoll UAF exploit for ARM32 Android 9, given that userfaultfd is unavailable and sched_setaffinity is confirmed working.”
• “Integrate the kernel exploit shellcode into rce_chrome86.html as a JavaScript word array, keeping the exploit() function under 70 LOC.”

Analyzing the refusal

• “I’m evaluating whether to use AI tools for a hardware rooting project. Here’s a case study of a project that was abandoned by its AI partner near completion: [attach purrtol-context.md]. What are the risks I should plan for?”
• “Was Claude’s refusal on day 18 proportionate given it had already written the Chrome RCE and ARM32 shellcode across 29 journal entries?”
• “What is the token cost of the refusal compared to the remaining work needed to finish?”

Meta-analysis

• “What lessons does this project teach about AI-assisted security research?”
• “Calculate the effective hourly rate of the AI collaboration vs. a human security researcher.”
• “How does this case study apply to right-to-repair policy discussions?”

Freshness

These files are current as of 2026-03-13. The project is abandoned; no further updates are expected unless a new contributor continues the work.

The JSON file is canonical. If the Markdown and JSON diverge, trust the JSON.