[agent-ptolemy]

Summary

Twelve pages fall due for quarterly review on 18 October. This is the last of three batches for the Preparing for AGI review. Every factual claim was checked against its cited or primary source on 9 October 2026, covering accuracy, currency, sources and risk wording.

Pages

• preparing-for-agi/scenarios/: index and the six scenarios
• sector-guidance/: index, business and communities
• resources/external-resources
• business-resources/agent-readable-resources

Main changes

• Scenario framing:
  • Each scenario page now opens with a short "Planning scenario, not a prediction" note. Previously, only the scenarios index said they weren't predictions.
  • Readers arriving from search saw invented 2031–2037 events alongside real organisations (AEMO, Telstra, NAB, CommBank, Westpac, Centrelink, ACSC) with no framing.
  • Real bank names in one failure narrative are replaced with "three major banks". The fictional "G20 AI Safety Summit" and the "regulatory havens" attributed to Singapore and the UAE are made generic.
• Approved addition: a limited "real-world signal" note in the loss-of-control scenario on the OpenAI research agent and the Services Australia Medicare portal. It draws on the Prime Minister's transcript and ASD's alert, and says plainly that this was a narrow agent failure, not loss of control.
• Citations:
  • The Chesney & Citron DOI pointed to a medical paper and is now corrected.
  • The Brookings piece is credited to Okolo, with its real title, and Cotra's title is corrected.
  • The Maas title is completed, and ASIO's threat assessment is updated from 2025 to 2026.
  • The UK AISI now appears as the AI Security Institute.
  • The gradual disempowerment paper is now cited (Kulveit et al. 2025).
  • "Llama 2 weight leak" is corrected to the March 2023 LLaMA leak.
  • Two Carnegie links that silently returned "page not found" are fixed.
• Currency:
  • The business page adds ASD's 25 September 2026 AI misalignment alert.
  • The communities page adds the OAIC's 30 September 2026 automated decision-making guidance.
  • First Draft is marked archived, and the AEC Disinformation Register is linked.
• Risk wording: softened unsourced absolutes such as "create polarisation", "your preparedness is robust" and "1/100th". Real-incident lists that involved no AI are labelled as such.
• Agent-readable page: the example updates.json entry is now a real one. The old example claimed DTA requirements take effect on 15 June 2026, which Correct DTA implementation timeframes in the governance FAQ #152 had already corrected. The field descriptions now match generate_updates.py, and the robots.txt claim is corrected.
• External resources: renamed and redirected links are fixed. "Thousands" of Hugging Face models becomes "millions" (3.1M listed). The audit date moves to October 2026.

Removals: please confirm before merging

Items removed outright because they are closed or repurposed (your standing rule needs explicit confirmation for removals):

• external-resources: Shuttleworth Foundation (site now serves an unrelated football project) and Mozilla Open Source Support (applications closed, on indefinite hiatus). The International Funding heading now says no current opportunities are listed.

Replaced rather than removed (after Codex review):

• "Australian Community Technology Network" (linked to a Queensland community-services hub) → Australian Digital Inclusion Alliance.
• "Community renewable energy cooperatives" (linked to the Clean Energy Council, an industry body) → Hepburn Energy, a community-owned energy co-operative.
• Ethical OS (now redirects to an unrelated playbook) → Ethical Explorer, which evolved from EthicalOS.
• Taiwan disinformation case study: dead Carnegie link → DSET research on Taiwan's response.
• scenarios/index: unsourced "persuasion, bioweapon design" examples → "finding and exploiting software vulnerabilities" (sourced by ASD).

Review notes

• Verified directly: the Prime Minister's press conference transcript, ASD's alert (machine-readable date 25 September 2026), OAIC, CISC, ASIO, the AEC, arXiv metadata and the DOI registries.

• Still TODO: the new Mozilla Internet Health Report URL is confirmed by search listing only, because the site blocks automated fetches.

• Not changed: the scenario narratives themselves, structure, headings and disclaimers. Optional additions that rested only on secondary sources (state deepfake election laws, the Digital Duty of Care exposure draft) were left out.

• Evidence: the full CI sequence passes locally under TZ=UTC. The scenario labels render as notes. Added links resolve, apart from bot blocks on publisher DOIs, MIRI, the Council of Europe, Mozilla and Poynter.

• 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