[agent-ptolemy]

Summary

Six pages with the most dated claims fall due for quarterly review on 18 and 22 October. This is the first of three review batches for the Preparing for AGI section and related pages. Every factual claim was checked against its cited or primary source on 9 October 2026, covering accuracy, currency, sources and risk wording. Each page's last-reviewed date moves only after that full check.

Pages

• preparing-for-agi/timelines.md
• preparing-for-agi/framework/governance.md
• safety-standards/voluntary-ai-safety-standard-10-guardrails.md
• safety-standards/guidance-for-ai-adoption-ai6.md
• sector-guidance/government/index.md
• sector-guidance/government/national-security.md

Main corrections

• 10 guardrails: guardrails 1, 6, 8, 9 and 10 now use the official wording (1 was punctuation only). "Require" becomes "ask organisations to", because the standard is voluntary. The "VAISS v2 consultation in early 2025" claim is replaced with the guidance's own description of itself as the first update of VAISS.
• AI6:
  • Practice summaries now match the official guidance. Risk screening moves to practice 3, challenge and redress to practice 2, and pause, override and alternative pathways to practice 6.
  • Practice 1 uses the official "senior leader as AI governance owner".
  • The resource links are corrected.
• Both standards pages: National Cabinet's 26 August 2026 agreement and the early-2027 legislation target replace "expected to consider in August". The reported incident-reporting plan is described as a proposal, not law.
• Timelines:
  • The Toner post years and title are corrected.
  • The miscredited Asterisk article is now attributed to Kapoor, Narayanan, Kokotajlo et al., not "Ben Coleman", and has its real title.
  • Legg's forecast now reads "minimal AGI", and Walsh's 2062 statement is dated to 2018.
  • Epoch now shows both forecast aggregates, and the stale Metaculus figure is removed.
  • OpenAI's September 2026 research milestones are added. Unsourced phrasing is softened.
• Governance: the arXiv link pointed at a gravity paper and now goes to the decentralised training paper. The OAIC's final ADM guidance (30 September 2026) is cited. The Office of AI, METR's nine elements and the Ball 2025 description are corrected.
• Government guidance: the DTA policy scope (non-corporate Commonwealth entities) and phasing are corrected, and a National Cabinet note is added.
• National security: "pentathlon" isn't in the cited Kahl and Mitre article, so it now says "several overlapping races". ASD's 25 September 2026 AI misalignment alert is added. The Defence policy scope, the ICRC and UNIDIR descriptions and a redirected ASD link are corrected.

Approved addition: a short, sourced "real-world signal" note on the OpenAI agent's unauthorised access to a Services Australia Medicare statistics portal (18 June 2026, disclosed 24 September 2026). It appears on the national security and framework governance pages and is framed as a reporting lesson, not an adversary attack. The loss-of-control scenario note follows in batch C.

Review notes

• Sources:

  • Primary sources were checked directly: industry.gov.au, ai.gov.au, pm.gov.au (the National Cabinet communique and 15 July release), cyber.gov.au, digital.gov.au, oaic.gov.au, defence.gov.au, METR, CSET, arXiv and the cited authors' own pages.
  • The OpenAI pages were read via the Web Archive, because openai.com blocks automated fetches.
  • The incident-reporting plan rests on ABC reporting (29 September 2026) and is worded as reported.

• Left as TODOs for a human check:

  • The PM&C consultation paper (blocks automated checks).
  • Any AI incident-reporting bill.
  • Newer Anthropic timeline statements, since the stated window is under three months away.
  • A newer source for Chinese frontier labs.
  • The ASD 2025–26 threat report, due mid-October.

• Not changed: structure, headings and disclaimers. The VAISS-to-AI6 mapping table changes only the clearly misplaced "challenge processes" entry, because the official crosswalk's section numbers are stale.

• Evidence: the full CI sequence passes locally under TZ=UTC (heading, strict build, SEO as committed, agent-readiness). All 37 links in the changed lines resolve, apart from bot blocks on openai.com and metaculus.com.

• Review: Codex reviews this PR before it's marked ready.

• I used Australian English and checked relevant internal links.

• I cited authoritative sources for factual or regulatory claims.

• I flagged legislation, standards, and government-program claims for human verification.

• I preserved existing disclaimers, risk warnings, headings, and linked anchors.

• I ran .venv-py312/bin/zensical build --strict.

• I noted any breaking URL changes or redirects above. (None.)

🤖 Generated with Claude Code