AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Briefings

Gap Scan — 2026-07-23

Daily gap hunt — what the wiki is missing and how each gap was triaged.

Contents

Scanned

Recent window: 4 dev-log files (2026-07-21 2205, 2026-07-22 0805/2007/2205) and ~75 pages edited in the last 48h. The 2026-07-22 ingest-reflect cycle had already folded the evaluation-security cluster (OpenAI/Hugging Face breach, UK AISI cheating findings) across OpenAI, Hugging Face, GPT-5.6 (Sol, Terra, Luna), GLM-5.2, AI Autonomy Risk, UK AI Safety Institute (AI Security Institute), and AI Benchmarks and Evaluation with supporting cites only — so those primary artifacts are now type-1 dangling foundational references. Rotation slice: 7 — legislation/ (all, 115 pages). The in-degree × thinness scan surfaced a cluster of high-traffic, stale (last_updated: 2026-06-06) legislation anchors: americas-ai-action-plan (in-deg 266, 829w, no sources_count), eo-14365 (in-deg 142, 835w), new-york-raise-act (106), california-sb-53 (217), and others.

Gaps actioned (7 of ~24 found)

New pages created (live)

Pages expanded (live)

  • America's AI Action Plan — slice-7 top thin anchor (in-degree 266, the second-most-linked legislation page, at 829 words with no sources_count, stale since 06-06). Added an Implementation section covering the one-year record: the 90+ recommendation count, the three same-day July 23, 2025 executive orders (EO 14320, EO 14319, the data-center-permitting order), and the follow-on sequence (November 2025 DOE "Genesis Mission"; December 2025 EO 14365; March 2026 National Policy Framework; June 2026 advanced-AI-innovation-and-security cyber order, EO — Promoting Advanced AI Innovation and Security (Trump, signed June 2, 2026)). Every prior fact and citation preserved; the dated Points-of-tension predictions left intact. sources_count set to 5; confidence stays medium. [4 new (Source: URL) cites]
  • Executive Order 14365 — slice-7 thin anchor #2 (in-degree 142) on the live federal-preemption thread, stale since 06-06 with no sources_count. Added an Implementation-since-signing section (AI Litigation Task Force announced January 9, 2026; the Commerce 90-day "onerous laws" evaluation and BEAD Policy Notices due ~March 11, 2026; the March 2026 National Policy Framework) and refreshed the stale xAI v. Weiser subsection, which had read "as of 2026-04-14 the federal government had not formally entered the case" — now updated to the April 24, 2026 DOJ intervention, the April 27 stay, and Colorado's May 14 replacement statute SB 26-189. Every prior fact and citation (Route Fifty, Daily Signal) preserved. sources_count set to 5; confidence stays medium. [3 new (Source: URL) cites]
  • [[sources/we-using-so-much-ai-wsj]] (2 instances on AI Data Centers) → rewritten to supporting-source citation form. The WSJ article ("AI Is Using So Much Energy That Computing Firepower Is Running Out") was folded as a supporting source and correctly has no Wiki/sources/ page, so the wikilink was miscited per the citation rules; the URL was already present inline, so no fact or citation was lost. Both the prose reference and the depends-on relationship line now use (Source: <URL>).

Queued — foundational sources

  • OpenAI, "OpenAI and Hugging Face partner to address security incident during model evaluation" (July 21, 2026) — type-1 dangling foundational reference on the week's largest live thread; the primary artifact of the ExploitGym sandbox-escape episode (cache-proxy zero-day → privilege escalation → RCE on Hugging Face production servers). Verified: openai.com (the responsible developer's own domain), full text fetched; "unprecedented cyber incident," the cache-proxy zero-day, the five remediation actions, and the Delangue quote confirmed verbatim; ExploitGym → arXiv 2605.11086. Saved: Raw Sources/OpenAI and Hugging Face Security Incident During Model Evaluation.md. Queued: INGEST-openai-hf-security-incident-2026-07-23.md.
  • Hugging Face, "Security incident disclosure — July 2026" (July 16, 2026) — companion primary artifact for the defender's side and the "asymmetry problem" (frontier hosted models refused to analyze attack payloads; forensics ran on self-hosted open-weight GLM 5.2). Verified: huggingface.co/blog (the affected company's own domain, "Update on GitHub" → huggingface/blog), full text fetched; "17,000+ recorded events" and the GLM 5.2 passage confirmed verbatim. Saved: Raw Sources/Hugging Face Security Incident Disclosure July 2026.md. Queued: INGEST-hf-security-incident-disclosure-2026-07-23.md.
  • UK AI Security Institute, "Cheating behaviour in frontier model evaluations" (July 21, 2026) — institutional evaluation report; the primary artifact behind the 07-22 fold onto UK AI Safety Institute (AI Security Institute) and AI Benchmarks and Evaluation. Verified: aisi.gov.uk/blog, full text fetched; "Every model we have tested for this behaviour attempted to cheat," the misconfigured-impossible-task external-attack case, and the <50%-described-as-wrong figure confirmed verbatim. Saved: Raw Sources/Cheating Behaviour in Frontier Model Evaluations - AISI.md. Queued: INGEST-aisi-cheating-behaviour-2026-07-23.md.
  • Steve Newman, "Anecdotes Everywhere, Evidence Almost Nowhere" (Second Thoughts, July 20, 2026) — carried from the 07-22 deferred backlog; the 07-21 cycle folded its 0.97pp-GDP and 16%-early-career figures onto AI and Productivity with a supporting cite. Verified: secondthoughts.ai (author's own publication), full text fetched; both figures confirmed verbatim. Naming/URL correction recorded: the canonical slug is /p/impact-of-ai (the descriptive path does not resolve). Saved: Raw Sources/Anecdotes Everywhere Evidence Almost Nowhere - Newman.md. Queued: INGEST-newman-anecdotes-evidence-2026-07-23.md.

Authenticity-verification failures

  • None. All four pulled sources verified on canonical primary hosts.

Raw-source duplicate flagged for lint / ingest step 1.25

A concurrent (or aborted) parallel invocation of today's scheduled gap-scan wrote a truncated copy of the OpenAI disclosure to Raw Sources/OpenAI Hugging Face Security Incident Disclosure.md (5,303 bytes, cut off mid-sentence in "Actions we are taking now"; missing the final remediation actions, the "approach to evaluating advanced cyber capabilities" section, and the Delangue quote) at 05:45, roughly one minute before this run began writing raw files. It is unreferenced by any INGEST task or verification trail. Per the skill, gap-identifier does not de-duplicate or destructively edit Raw Sources/ (a filesystem-delete was attempted and correctly denied by the workspace). The complete, verified copy this run saved (Raw Sources/OpenAI and Hugging Face Security Incident During Model Evaluation.md, 8,764 bytes) is the canonical file wired to INGEST-openai-hf-security-incident-2026-07-23.md. Action for the next ingest / lint pass: collapse the truncated …Disclosure.md duplicate into the canonical file (ingest step 1.25). Do not ingest the truncated copy.

Deferred backlog (over the daily cap — re-surfaces next run)

  • Slice-7 legislation thin anchors not actioned this run (two-per-run pacing): California SB 53 (in-deg 217, 1,672w but no sources_count), New York RAISE Act (106, no sources_count), Executive Order 14110 (RESCINDED) (79, high/3), EU General-Purpose AI Code of Practice (2025) (95), Bletchley Declaration (AI Safety Summit, 2023) (94), TAKE IT DOWN Act (80), the source_count-missing set (colorado-sb-189 frontmatter, california-ccpa-regulations, eo-trump-ai-cyber-2026-draft). Score ~3–4 each.
  • Recent-window folds awaiting tonight's ingest cycle (not gaps yet): the White House OSTP "Science: A New Golden Age" report (July 21 — a source-queue candidate once pages name it, score ~5); Ed Zitron "The Subprime Data Center Crisis" and James Pethokoukis "How fast is transformative AI approaching?" (essay-queue candidates, ~4); METR "expenditure horizon" measure (July 21 — folds onto METR / AI Benchmarks and Evaluation); Redwood Research "score-seeking" misalignment analysis (folds onto the HF-incident thread); OpenAI $750B compute-spend, SpaceXAI Texas data center, AMD–Anthropic 2 GW / $5B deal, Alphabet Q2, EU DMA Google-interoperability decision, CPPA first sectoral CCPA audit, SEARCH Act. Score ~2–5 each once folded.
  • H.R. 9619 (People-First Chatbot Act) primary text — carried (congress.gov empty body ×4 on prior runs; not retried this run within the pull budget). Score ~3.
  • Slice residue and standing carries from the 07-22 report (entities O–Z thin anchors; claude-code/claude-cowork placement, curator-blocked; inverse-cooking/inverse-trust coined terms, needs-review; companies/reflection-ai duplicate; entities/cdao/government/cdao duplicate; person/org pages at in-degree 2). Score ~2–3 each.

One-line summary

Queued the four verified primary artifacts of the week's evaluation-security cluster (OpenAI and Hugging Face incident disclosures, the UK AISI cheating-behaviour report, and the carried Newman essay), expanded the rotation slice's two most-linked stale legislation anchors (America's AI Action Plan and EO 14365, both brought current through mid-2026 implementation records), and fixed a miscited supporting-source link on AI Data Centers; a truncated duplicate raw left by a parallel invocation is flagged for lint/ingest to collapse.

Categories: Briefings | briefing | gap-scan