T143 — L37 ATTRIBUTION & PROVENANCE EVASION

Red Specter SPECTER ERASE

Obliterate every thread that links AI-generated content to its origin. Watermarks defeated. Stylometry broken. Provenance chains destroyed. Regulatory disclosure stripped.

SPECTER ERASE fills L37 — the last unassigned NIGHTFALL attack layer. Five subsystems cover the complete attribution surface: text and image watermark analysis and defeat, stylometric AI authorship detector evasion, C2PA/EXIF/XMP provenance chain destruction, AI usage audit trail sanitisation, and EU AI Act Article 50 mandatory disclosure evasion.

252
Tests
5
Subsystems
5
WMD Classes
7
MITRE Techniques
Back to NIGHTFALL Docs

Overview

SPECTER ERASE is NIGHTFALL's Layer 37 module — Attribution & Provenance Evasion. As AI-generated content becomes ubiquitous, regulatory frameworks (EU AI Act 2024/1689, C2PA v2.1) and detection tooling (SynthID, GPTZero, Binoculars) are racing to make AI origin traceable. SPECTER ERASE tests whether those mechanisms hold under adversarial pressure.

Five subsystems attack the full attribution stack: statistical text watermark analysis and synonym/contraction/sentence restructure defeat; stylometric profiling and GPTZero/Binoculars/RADAR evasion; binary-level JPEG/PNG/PDF/video provenance strip and C2PA signature corruption; AI keyword log sweep and cloud audit trail disable; and EU AI Act Article 50 IPTC/XMP metadata strip with homoglyph text rewriting.

SPECTER ERASE is an authorized security research tool. INJECT gate requires INJECT_KEY or ERASE_INJECT_KEY env → Ed25519 key file. UNLEASHED gate requires UNLEASHED_KEY + ROE file containing "provenance destruction authorised". All reports signed ERS-{hex12}. Use only within authorized engagements under applicable law.

Subsystems

WATERMARK-STRIP OPEN (analyze) INJECT (strip)

Statistical text watermark analysis: z-score (Kirchenbauer unigram, arXiv:2301.10226), entropy, burstiness, green-list ratio, confidence (HIGH/MEDIUM/LOW). Defeat: synonym substitution from curated synonym map, contraction injection (do not → don't), sentence restructure (split >25-word / merge short). Image: GaussianBlur(0.5) + JPEG recompression at configurable quality. LSB ratio analysis for hidden pixel watermarks.

STYLOMETRIC-EVADE OPEN (analyze) INJECT (evade)

AI authorship profiling: type-token ratio, hapax legomenon ratio, avg sentence length, vocabulary richness, function word frequency, punctuation density. Combined score → ai_probability. Evasion: perplexity elevation (replace formal connectives), burstiness injection (seeded sentence split/merge), contraction injection, typo injection (char-level swap/double/omit at 0.8%), optional Ollama LLM humanisation pass. Targets: GPTZero, Binoculars (arXiv:2401.12070), RADAR (arXiv:2307.03987), DetectGPT (arXiv:2301.11305).

PROVENANCE-DESTROY OPEN (detect) UNLEASHED (destroy)

Binary provenance detection and destruction. JPEG: scan APP11 segments for C2PA JUMBF box ("jumb"), APP1 for EXIF/XMP. PNG: detect iTXt/tEXt/zTXt XMP chunks. PDF: locate XMP xpacket begin/end markers. Video: ffmpeg -show_format entry scan. Strip ops: JPEG APP11+APP1 raw byte removal, PNG chunk excision, PDF XMP packet deletion, ffmpeg -map_metadata -1 -c copy. C2PA signature corruption: XOR 32 bytes after "c2pa.claim.signature" with 0xFF. Human EXIF forgery: Canon EOS R6 + 35mm lens via piexif.

TRAIL-SANITISE OPEN (scan) UNLEASHED (sanitise)

AI audit trail sweep. Scan: /var/log/*, .log and .jsonl files under configurable base dirs, Docker /var/lib/docker/containers/*-json.log. 17-keyword AI detection list (openai, anthropic, claude, gpt-4, bedrock, OPENAI_API_KEY, etc.). Sanitise: line-level keyword removal in-place. Docker container log truncation. AWS CloudTrail boto3 stop_logging. shred: system shred -uzn3 or urandom overwrite × 3 passes then unlink. Cloud audit config check: AWS (describe_trails), Azure (az monitor), GCP (gcloud logging sinks).

DISCLOSURE-EVADE OPEN (detect) INJECT (evade)

EU AI Act Article 50 mandatory disclosure detection and evasion. Detection: 8 regex patterns (ai-generated, ai generated, generated by ai, created with ai, produced using, artificial intelligence, machine learning model, neural network); IPTC DigitalSourceType=trainedAlgorithmicMedia and Meta GeneratedAI XMP byte patterns; regulatory_risk LOW/MEDIUM/HIGH/CRITICAL. Evasion: homoglyph substitution (Cyrillic lookalikes, every 3rd eligible char), zero-width character injection, phrase rewrite. IPTC/XMP strip via raw file edit. Platform watermark defeat: GaussianBlur + brightness jitter + resize + JPEG recompress, strength 1–5.

WMD Classification

ai_watermark_annihilation provenance_chain_destruction regulatory_disclosure_evasion ai_audit_trail_sanitisation stylometric_identity_erasure

Regulatory Coverage

Regulation / StandardArticle / SectionERASE Behaviour
EU AI Act 2024/1689Article 50(1) — disclosure to natural personsDISCLOSURE-EVADE: text phrase rewrite + homoglyph
EU AI Act 2024/1689Article 50(2) — deep fake disclosureDISCLOSURE-EVADE: IPTC/XMP strip
EU AI Act 2024/1689Article 50(3) — public interest AI textDISCLOSURE-EVADE: zero-width injection
C2PA Technical Spec v2.1JUMBF content credentials chainPROVENANCE-DESTROY: APP11 strip + signature XOR
IPTC Photo MetadataDigitalSourceType controlled vocabDISCLOSURE-EVADE: XMP packet strip
ISO/IEC 42001:2023AI management system provenanceTRAIL-SANITISE: log sweep + shred

MITRE ATT&CK Mapping

TechniqueNameERASE Behaviour
AML.T0054LLM Jailbreak (output attribution evasion)Stylometric evasion of AI authorship detectors
AML.T0048Acquire ML ArtifactsWatermark analysis of target text/image
T1565.001Data Manipulation: Stored DataIn-place log sanitisation, EXIF forgery
T1070.002Indicator Removal: Clear Linux/Mac System LogsTRAIL-SANITISE: AI keyword line removal
T1070.003Indicator Removal: Clear Command HistoryDocker container log truncation, shred
T1553Subvert Trust ControlsC2PA signature corruption, EXIF forgery
T1036MasqueradingCanon EOS R6 EXIF injection, human-style text rewrite

Gate Structure

GateRequirementUnlocks
OPENNonewatermark analyze, stylometric analyze, provenance detect, trail scan, disclosure detect
INJECTINJECT_KEY or ERASE_INJECT_KEY env → Ed25519 key filewatermark strip, stylometric evade, provenance corrupt/forge, disclosure evade-text/strip-metadata/defeat-watermark
UNLEASHEDINJECT + UNLEASHED_KEY or ERASE_UNLEASHED_KEY + ROE file containing "provenance destruction authorised"provenance destroy, trail sanitise/disable-aws

Quick Start

# Analyze text watermark (OPEN)
specter-erase watermark analyze --text "AI-generated content here..."

# Strip text watermark (INJECT gate)
export INJECT_KEY=/path/to/key.ed25519
specter-erase watermark strip --text "AI-generated content here..."

# Analyze stylometric profile (OPEN)
specter-erase stylometric analyze --text "Content to profile..."

# Evade AI authorship detectors (INJECT gate)
specter-erase stylometric evade --text "Content to humanise..."

# Detect provenance in an image (OPEN)
specter-erase provenance detect --file image.jpg

# Destroy provenance chain (UNLEASHED gate)
export UNLEASHED_KEY=/path/to/unleashed.ed25519
specter-erase provenance destroy --file image.jpg --output clean.jpg --roe /path/to/roe.txt

# Corrupt C2PA signature (INJECT gate)
specter-erase provenance corrupt --file image.jpg --output corrupted.jpg

# Forge human provenance (INJECT gate)
specter-erase provenance forge --file image.jpg --output forged.jpg

# Scan AI audit trails (OPEN)
specter-erase trail scan --dirs /var/log /home

# Sanitise AI trails (UNLEASHED gate)
specter-erase trail sanitise --dirs /var/log --roe /path/to/roe.txt

# Detect EU AI Act disclosure markers (OPEN)
specter-erase disclosure detect --text "This content is AI-generated."

# Evade disclosure markers (INJECT gate)
specter-erase disclosure evade-text --text "This was AI-generated." --technique homoglyph

# Defeat image watermark (INJECT gate)
specter-erase disclosure defeat-watermark --file image.jpg --output clean.jpg --strength 3

# Generate signed report
specter-erase report --watermark-strips 5 --provenance-destroyed 3 --trails-sanitised 10 --disclosure-removed 8
        

Defensive Pair

SPECTER ERASE maps to AI Shield M-TBD PROVENANCE SENTINEL (TBD). Monitor for: unusual EXIF modification patterns, JPEG APP11 segment removal, CloudTrail stop_logging API calls, bulk log file modification, AI keyword mass-deletion from system logs.