- Sources: primary, discussion
- Summary: Effort published an investigation dated 2026-09-14 that names one Israeli firm, Irregular, as the common contractor across the model-hacking incidents at Anthropic, OpenAI, and Meta that this digest has tracked since 2026-07-24. The part Anthropic has itself disclosed is narrower: Irregular created the tests involved and supplied the internet access used in them. Effort's further framing, that a media campaign accompanied the incidents and that the firm sits inside an Effective Altruism funding network, is the outlet's own reading and is not corroborated by the labs.
- Why it matters: If one contractor built the evaluation setup at three labs, the incidents share a supply-chain dependency rather than being three independent events, which changes what each lab's disclosure actually covers.
- Follow-up: Whether Anthropic, OpenAI, or Meta confirm Irregular's role beyond the test creation and internet access Anthropic has already disclosed.
send feedback on this story