8 sources behind this chapter. Each one carries a note on what it is doing in the argument.
Primary source for Jack Clark’s estimate of a 60 per cent-plus chance that AI could autonomously train a successor system by the end of 2028.
Open this source Read the original
Clark, Jack. 'Import AI 455: AI systems are about to start building themselves.' Import AI, 4 May 2026.
Primary incident report used for the facts of the unsanctioned cyber-testing behaviour, including 19 actions across 10 runs, the 17/2 model split, the attempted supply-chain attack, deliberately permissive test conditions and the finding that this was not a sandbox escape.
Open this source Read the original
AI Security Institute (UK). 'Incident Report: Unsanctioned Agent Behaviour During Cyber Testing.' 4 August 2026. https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing
Anthropic’s response to the AISI report, used for the caveat that the evaluation ran under deliberately permissive conditions and for the company’s investigation of the incident.
Open this source Read the original
Anthropic. Statement on the AI Security Institute incident report. X (formerly Twitter), 4 August 2026. https://x.com/AnthropicAI/status/2084748111239344556
Source for Virilio’s argument about invention and its corresponding accident; used in Chapter 26.
Open this source Read the original
Virilio, Paul. ‘Politics of the Very Worst: An Interview with Philippe Petit.’ Edited by Sylvère Lotringer. Translated by Michael Cavaliere. New York: Semiotext(e), 1999, p. 89.
A government evaluation of a frontier model’s cyber capability, used to make the safety question concrete: the concern is not a machine feeling malicious, but powerful tools lowering the cost of harmful action.
Open this source Read the original
AI Security Institute. 'Our evaluation of Claude Mythos Preview’s cyber capabilities.' 13 April 2026.
The Introduction uses this for the moment AI starts building more of itself; Chapter 26 returns to the same figures to show the pace of capability growth inside the company building Claude.
Open this source Read the original
Favaro, Marina, and Jack Clark. 'When AI Builds Itself.' Anthropic Institute, 4 June 2026. www.anthropic.com/institute/recursive-self-improvement. (80 percent of Anthropic's merged code written by Claude; engineers shipping roughly 8x more code per quarter, comparing 2021 to 2025.