Project Simurgh
Project Simurgh is the defensive counterpart to The Invisible Window research: a zero-trust integrity API for autonomous agents and high-stakes proctoring. Instead of trusting a visual stream that can be structurally bypassed, Simurgh validates behavioral and environment metadata, builds tamper-evident audit records, and keeps the integrity signal privacy-preserving.
01. Problem
The Invisible Window shows that browser and OS screen-capture pipelines cannot be treated as ground truth. Proctoring platforms and agentic AI systems that rely on screenshots or UI vision can be deceived by documented display-affinity APIs and click-through overlays. A safer integrity layer needs to verify behavior and environment state without expanding surveillance.
02. Solution Overview
- Built a metadata-only integrity pipeline that evaluates behavioral telemetry rather than screen contents
- Added Academic Shield flows for exam creation, session join, privacy acceptance, telemetry submission, review reports, and audit verification
- Separated deterministic local scoring from optional AI narrative analysis so provider failures do not break the authoritative score
- Added native-helper and Local Integrity Node direction for detecting display-affinity and producing signed proof envelopes
- Anchored the project to The Invisible Window threat model while keeping the implementation vendor-neutral and privacy-preserving
Build
Tech Stack
- • Samples lightweight behavioral telemetry windows instead of recording screen pixels
- • Detects focus loss, bulk paste, idle gaps, typing anomalies, and display-affinity risk signals
- • Stage 1 Academic Shield workflow covers exam lifecycle, privacy acceptance, local risk scoring, reports, and audit verification
- • Stage 2 proof pipeline signs privacy-preserving integrity envelopes for future device-level trust
Secure
- No screen pixels, webcam frames, audio, typed answer content, pasted content, or personal identity data collected
- Student identifiers hashed before storage
- Instructor, helper, audit, and session boundaries separated with dedicated secrets
- Replay protection rejects duplicate sequences, stale timestamps, future timestamps, and malformed telemetry
- HMAC-SHA256 linked audit chain makes report tampering detectable
- Privacy-preserving reviewer model: Simurgh produces review recommendations, not automatic misconduct findings
03. Proof & Verification
Verified Claims
- >Archived with DOI 10.5281/ZENODO.20195198
- >Stage 1 research MVP and Stage 1.5 validation pack documented in the repository
- >Telemetry payloads are lightweight behavioral JSON windows, not video streams
- >Audit verification endpoint validates the HMAC-linked event chain
- >GitHub Stage 1 checks run the project quality gate on main and pull requests