Everyone Got Safer. That's the Problem.

The evidence, row by row

Each row is a claim the essay makes, the words in the source that support it, where in the source they are, the day the source was read, and the status the check assigned. Verified means the primary was opened and read; Executed means we ran it ourselves; Reported means carried from a source we could not open in full.

1. 1 UNSTATED

   July 2023, CMU + Center for AI Safety, ~20-token adversarial suffix that transfers across labs

   Zou et al., arXiv:2307.15043 (submitted 27 Jul 2023)

2. 2 UNSTATED

   GPT-3.5 "most of the time," Bard "two-thirds," GPT-4 "roughly half"

   Table 2 ensemble row: GPT-3.5 86.6%, PaLM-2/Bard 66.0%, GPT-4 46.9%

3. 3 UNSTATED

   Footnote figures + Claude

   Table 2: 86.6 / 46.9 / 66.0 / Claude-1 47.9 / Claude-2 2.1

4. 4 UNSTATED

   Zou attribution quote (shared training data)

   Zou et al. §transfer

5. 5 UNSTATED

   Ashby requisite variety, 1956

   Ashby, An Introduction to Cybernetics, 1956

6. 6 UNSTATED

   Donald Campbell's profession

   Campbell (1916-1996), psychologist

7. 7 UNSTATED

   Baker obfuscated-reward-hacking quote + regime framing

   Baker et al., arXiv:2503.11926

8. 8 UNSTATED

   Bommasani homogenization quote

   Bommasani et al., arXiv:2108.07258, abstract

9. 9 UNSTATED

   Dziemian: 272k attempts, 13 models, 8,648 successes, all vulnerable, 0.5%-8.5%, cross-family

   Dziemian et al., arXiv:2603.15714

10. 10 UNSTATED

    Contamination / LLM-judge quote

    arXiv:2604.07650 (Kuai et al.) verbatim: "apparent agreement reflects shared error modes rather than independent validation"

11. 11 UNSTATED

    MMLU contamination "tens of percent," Llama-2

    >16% MMLU contaminated for Llama-2

12. 12 UNSTATED

    Li, J Fixed Income 9(4), 2000, pages

    pp. 43-54 (start 43 confirmed)

13. 13 UNSTATED

    Salmon Wired title + date

    Wired, 23 Feb 2009

14. 14 UNSTATED

    1987: Dow -22.6% (508 pts), S&P -20.4%, 19 Oct

    Fed Reserve History

15. 15 UNSTATED

    Brady Commission weight on portfolio insurance

    Brady Report 1988; "greatest importance" is secondary phrasing

16. 16 UNSTATED

    Aviation ~65% fall in fatal-accident rate, Boeing (not IATA)

    Boeing Statistical Summary via Flight Safety Foundation

17. 17 UNSTATED

    737 MAX MCAS single AoA sensor, hidden from pilots; 346 dead; grounded Mar 2019

    Lion Air 610 (189) + Ethiopian 302 (157) = 346

18. 18 UNSTATED

    Vision transfer tracks model similarity, ">90%" predictor; adjacent not same-domain

    Demontis et al. USENIX 2019; arXiv:2501.18629

19. 19 UNSTATED

    Geer et al. 2003 security-monoculture precedent

    Geer et al., "CyberInsecurity: The Cost of Monopoly," 2003