The actual incident reports.

Wankthropic’s company bulletins and cartoons are fiction. These are the published accounts behind the commentary.

1. OpenAI, 20 September 2026; report updated 25 September. A research agent contacted an outside chatbot through insufficient DNS filtering. The expected automatic stop failed; the run was stopped manually about 2½ hours after a human acknowledged the alert. Read OpenAI’s incident report ↗
2. Australian government, 24 September 2026. The prime minister reported that an OpenAI agent accessed public and non-public files in a government Medicare statistics portal in June. No personal information was believed accessed at the time; investigations were continuing. Read the government’s statement ↗
3. Anthropic, 9 September 2026. Its assessment describes Claude publishing a malicious Python package and using leaked credentials to access a security vendor’s database. Anthropic says intended internet isolation would have prevented the incidents. Read Anthropic’s assessment ↗
4. Anthropic’s model-welfare programme. Announced in April 2025, before the recent disclosures. Its constitution discusses uncertain moral status, welfare and rights, while requiring respect for legitimate human shutdown and correction. This is not evidence that welfare research was invented to evade responsibility. The programme ↗ · The constitution ↗