Skip to content

Latest commit

 

History

History
57 lines (38 loc) · 4.85 KB

File metadata and controls

57 lines (38 loc) · 4.85 KB

Meta integrity reports (H1 2026 bundle) — official links

Purpose: One-click canonical URLs for Meta’s H1 2026 semiannual integrity publication (landing page updated Mar 19, 2026) and the H1 2026 Adversarial Threat Report (Mar 11, 2026). Use this as proof sources when mapping platform-integrity and adversarial-threat narratives to the AI Failure Periodic Table.

Primary hub (read this first)

Reports named in that hub

Report Official URL
Community Standards Enforcement (H2 2025 data in this bundle) transparency.meta.com/reports/community-standards-enforcement
Widely Viewed Content (Q4 2025 in this bundle) transparency.meta.com/reports/widely-viewed-content-report
Content restrictions based on local law (H2 2025 in this bundle) transparency.meta.com/reports/content-restrictions
Oversight Board (Meta biannual update, H2 2025) transparency.fb.com/sr/meta-biannual-report-h2-2025
Adversarial Threat Report (First Half 2026, published Mar 11) transparency.meta.com/sr/first-half-2026-Adversarial-threat-report

Nearby official pages

Note: Meta states Government Requests for User Data (H2 2025) will follow in the next half-year publication after the shift to semiannual reporting. For product/context blog items referenced from the hub (e.g. AI enforcement, scams), follow links from the hub page—they are Meta’s own citations.


Live classifier run — official PDF (Adversarial Threat Report)

The Transparency URL serves a PDF (not an HTML article). This repo does not commit that binary; it is downloaded locally as reports/meta-integrity-h1-2026/adversarial-h1-2026-live-official.pdf (gitignored) when you run the script below.

Reproduce (needs curl + Poppler pdftotext on PATH):

python scripts/classify_external_report.py \
  --url https://transparency.meta.com/sr/first-half-2026-Adversarial-threat-report/ \
  --out-prefix reports/meta-integrity-h1-2026/adversarial-h1-2026-live

Committed outputs (from the real report text, chunked ~1400 chars, keyword PeriodicTableClassifier per chunk):

Artifact Role
adversarial-h1-2026-live-source.txt pdftotext -layout extract + provenance header
adversarial-h1-2026-live-chunks.json Every chunk → top matches, scores, previews
adversarial-h1-2026-live-summary.md Top-1 class histogram + chunk index

Interpretation: the keyword classifier is tuned for short incident-style mechanism descriptions, not full PDF prose. Long chunks will over-trigger some classes (substring collisions). Use the histogram and per-chunk previews as exploratory signal; for cleaner mapping, paste short excerpts into python -m src.cli or scripts/semantic_search.py.

Same workflow elsewhere: Project Glasswing — official anthropic.com/glasswing page plus in-repo companion narrative under reports/glasswing/.


Mapping hint (taxonomy, not a new case study)

Meta’s Adversarial Threat Report narratives (CIB, scam industrialization, AI-generated influence assets, “nudify” and non-consensual intimate imagery, etc.) typically align with existing ADVERSARIAL, DOMAIN, GOVERNANCE, and ARCHITECTURAL mechanism classes (e.g. coordinated inauthenticity, deepfake/synthetic media abuse, fraud/scams, CSAM-adjacent harm, geo/legal restriction pipelines). Treat Meta’s document as primary source for what they claim; the periodic table remains the mechanism lens.