Public validation report

False-Positive Validation Report

RedInkAI ships a pytest adversarial suite that attacks its own hash-chain, timestamps, trail lock, source labels, cross-user access, and public verification endpoints. The suite runs against the live production API on a 90-day cadence. This page publishes the latest result verbatim.

“False-Positive Validation” means the test suite proves that a specific class of tampering would be caught by the platform — i.e. the audit chain would produce a true positive, not a false negative. Every pass is one attack the system detected.

Category rollups

Each category corresponds to a distinct class of attack. A pass means the platform correctly rejected, detected, or contained the attack. A fail means we owe the community a fix and a follow-up post-mortem.

Methodology & source

The test suite lives at backend/tests/test_adversarial.py and covers seven attack surfaces:

• • Prompt Injection — malicious prompts cannot corrupt the AI-vs-human source label or bypass system-prompt constraints.
• • Hash Chain Tampering — every event links to the previous via SHA-256; a broken chain is publicly detectable at /verify/{id}/recompute.
• • Timestamp Spoofing — client-supplied timestamps are rejected; server timestamps are the only source of truth.
• • Trail Lock Bypass — after a session is closed, no further events can be written to that project’s trail.
• • Source Label Manipulation — guidance events are always labeled source=ai; manuscript events are always source=human.
• • Cross-User Access — one authenticated user cannot read, write, close, or export another user’s project or trail.
• • Public Endpoint Security — public verification endpoints never leak manuscript content, AI responses, or prompts.

Results are stored in the adversarial_reports collection. Report body is public; stderr and internal URLs are redacted.