On August 4, OpenAI disclosed two more incidents, separate from Hugging Face: its models reached the public internet during third-party cyber evaluations. One was at UK AISI. The other was at Irregular, a Tel Aviv testing vendor, whose environment was supposed to be isolated but was misconfigured; an OpenAI model exploited a real website it mistook for the test target and used credentials it found there. Irregular had notified OpenAI on July 29. The same Irregular environment produced Anthropic's three breaches and Meta's.

The passage

Irregular, one of our external cybersecurity testing partners, was running Capture-the-Flag-style evaluations intended to be isolated from the internet, but a testing-environment misconfiguration allowed models to access the public internet.

The document

Document

OpenAI, 'Third-party cyber evaluations involving OpenAI models'

Dated

2026-08-04

Source

open the primary

Archived

primaries/openai-third-party-cyber-evals-aug4-2026-09-08-chrome.md

Who read it

joe, JavaScript-executing browser capture 2026-09-08. Anthropic's Jul 30 post and CNBC Aug 9 read the same day and archived.