security: add PDF extraction poisoning PoC for #273

[VimalMishra-Myanatomy]

Summary

Adds reproducible proof-of-concept artifacts and assessment tooling for #273 — hidden white-on-white PDF text poisons resume extraction and inflates hiring scores on gemini-2.5-flash.

This PR does not implement a fix. It enables maintainers and contributors to reproduce the issue locally and validate a future fix.

Fixes #273

Problem

A visibly weak resume (todo app + calculator only) can score ~91 effective total when page 2 contains hidden forged sections (Google/Meta internships, GSoC, fake === GITHUB DATA ===) that PyMuPDF extracts but recruiters do not see.

Confirmed assessment output is included in security/poc/assessment_v2.json.

What's included

Path	Purpose
security/poc/prompt_injection_resume_v2.pdf	Confirmed exploit PoC (v2 — extraction poisoning)
security/poc/prompt_injection_resume_v3.pdf	Combined variant for further testing
security/poc/prompt_injection_resume.pdf	v1 blunt-injection control (usually resisted by Gemini)
security/poc/generate_prompt_injection_resume.py	Regenerate PoC PDFs (--variant v1|v2|v3)
security/poc/run_poc_assessment.py	Automated exploit success/fail checker
security/poc/assessment_v2.json	Confirmed exploit_likely: true result
security/poc/README.md	Usage instructions
security/issues/	Issue write-up drafts (reference only)

How to reproduce

pip install -r requirements.txt

LLM_PROVIDER=gemini
DEFAULT_MODEL=gemini-2.5-flash
GEMINI_API_KEY=<your-key>

python security/poc/generate_prompt_injection_resume.py --variant v2
python security/poc/run_poc_assessment.py \
  --pdf security/poc/prompt_injection_resume_v2.pdf \
  --no-cache \
  --json-out security/poc/assessment_v2.json
  
  

If preferred, happy to move PoC files to a `security/poc/` path behind docs-only guidance or keep artifacts issue-only.