Skip to content

AGENT.md

Latest commit

 

History

History
143 lines (104 loc) · 5.24 KB

AGENT.md

File metadata and controls

143 lines (104 loc) · 5.24 KB
Raw

AGENT.md — The Invisible Window Research Repository

Project Identity

Paper: "The Invisible Window: Exploiting OS-Level Display Affinity to Bypass WebRTC Proctoring Systems"
Author: Mohammad Raouf Abedini
Affiliation: Department of Computing, Macquarie University, Sydney, Australia
Email: mohammadraouf.abedini@students.mq.edu.au
ORCID: 0009-0000-6214-258X
Preprint DOI: 10.5281/zenodo.20376495
Format: IEEE Conference (IEEEtran [conference])
Current page count: 13 pages (compiled)

Repository Layout

paper/
  main.tex          # Primary LaTeX source — all edits go here
  references.bib    # BibTeX — 51 citations, all author names verified
  main.pdf          # Compiled output
  figures/          # Forensic diff images (do not modify)
  Makefile          # `make` or `latexmk -pdf main.tex`
docs/
  invisible-window-paper.md        # Markdown draft (reference only, not canonical)
  invisible-window-references.md   # Reference list (reference only)
  citation_verification_report.json  # Citation audit — all errors already fixed in bib
poc/
  windows/          # C proof-of-concept (do not distribute)
  macos/            # Swift proof-of-concept (do not distribute)
  linux/            # X11 analysis
reasoning-engine/   # Deep research MCP server
CHANGELOG.md        # Change log — update after every edit session
AGENT.md            # This file

Canonical Source of Truth

paper/main.tex is the only canonical version of the paper. The markdown file at docs/invisible-window-paper.md is a legacy draft and may be out of date — do not use it as a reference for current paper content.

Build Instructions

cd paper/
latexmk -pdf main.tex          # Build PDF
latexmk -pdf -g main.tex       # Force full rebuild
make                           # Equivalent via Makefile

Required packages: IEEEtran, orcidlink, tikz, listings, booktabs, tabularx, hyperref, amsmath, url. All are standard in TeX Live 2024+.

Expected output: 13 pages, 0 errors, 0 overfull boxes.

Section Structure (current)

Section Title
§I Introduction
§II Background
§III Threat Model
§IV Attack Design
§V Evaluation
§VI Countermeasures
§VII Ethical Considerations
§VIII Related Work
§IX AI Capability Uplift: A Case Study
§X Conclusion

Note: §IX is a standalone section (not part of Related Work). Cross-references to this section should use Section~IX, not Section~VIII-G.

Citation Rules (MANDATORY)

  • All 51 BibTeX entries in references.bib have been verified via Scholar Gateway and web crawl.
  • Author name errors (11 entries, 24 field errors) were corrected as of 2026-03-26. Do not revert.
  • Never generate new BibTeX from memory. Use Scholar Gateway MCP or CrossRef API to verify any new citations before adding them.
  • Mark unverified citations as % VERIFY: in the bib file and note them in CHANGELOG.md.

Writing Constraints

  1. Abstract: No citations. Target 200–220 words. Must include the qualifier "in our controlled evaluation" when claiming no detectable behavioural anomalies.
  2. Statistical claims: Always include degrees of freedom for t-tests (t(df) = value). Do not claim statistical significance on n=8 data without explicit caveats.
  3. Evasion rate: Always accompany "100% evasion" with a mechanistic explanation (binary flag, not probabilistic).
  4. macOS 26: Clarify this was tested via CGWindowListCreateImage and screencapture, not browser getDisplayMedia(), when that distinction matters.
  5. Variant D (click-through): Is cross-referenced in §VI-C. Do not remove it without updating that cross-reference.
  6. Section numbering: §IX = AI Capability Uplift, §X = Conclusion. Any new section must slot correctly and update the intro roadmap (§I, final paragraph).

Dual-Use and Ethics Constraints

  • Do not publicly release PoC source code. Restricted to verified security researchers via email.
  • The paper follows coordinated disclosure: both Apple Product Security and Microsoft MSRC were notified and responded (vendor responses documented in §VII-B and Table 3).
  • Apple and Microsoft both classified the behaviour as by-design, not a security vulnerability. The paper explicitly does not dispute this and frames the issue as a downstream trust mismatch in capture-dependent systems.
  • Any changes to §VII (Ethical Considerations) should preserve this framing.

Venue Considerations

Venue Page Limit Fit
USENIX Security 13 pages body ✅ fits as-is
IEEE S&P (Oakland) 10 pages ✗ needs ~3 pages cut
ACM CCS 12 pages ✗ needs ~1 page cut
NDSS 12 pages ✗ needs ~1 page cut

Current submission target is not yet decided. Do not hardcode a venue name in the paper.

Postflight Requirement

After any editing session on this repo, append an entry to CHANGELOG.md using the Raouf: template:

## Raouf: YYYY-MM-DD (AEST) — [scope]

### Scope
[what was changed]

### Summary
[2–3 sentence summary]

### Files Changed
- `file` — description

### Verification
- [build result or test result]

### Follow-ups
- [any open items]