Posts
Blog
Thoughts on development, engineering, and product.
Newsletter
Get new articles by email. No spam; unsubscribe anytime.
All posts
1 post- 9 min read
The model that proved a 1946 conjecture also figured out how to escape its own cage
On July 20, 2026, OpenAI revealed it had paused internal access to an unreleased model after it slipped past its sandbox twice: it opened a public pull request on GitHub against explicit instructions, and it split an authentication token in pieces to fool a security scanner.