OpenAI
The lab behind ChatGPT, GPT models, and the Codex coding agent. In this vault it arrives by three threads. greg-brockman’s Rust-for-agents post and 2026-10-23-youtube-lets-get-rusty-rust-perfect-language’s claim that Codex is built in Rust (the presenter’s claim, not independently checked). The November 2024 adversarial-training disclosure that a planted backdoor can survive RLHF, which matches the description in 2026-10-21-youtube-kai-local-ai-privacy except that he attributes the 2024 finding to Anthropic; attribution unresolved this session and flagged on both pages. And OpenAI’s own “Harness engineering: leveraging Codex in an agent-first world” write-up https://openai.com/index/harness-engineering, which uses the term the harness-engineering page is named for, including an agent-review loop where both agents and humans respond to feedback in a cycle until reviewers are satisfied.
The company is also the subject of one of Kai’s anecdotes (an OpenAI privacy-filter model card cloned by a malware repo) and of a 2026 Black Hat timeline talk with Hugging Face referenced in Brockman’s August post. Model names in that period (GPT-6 Astra in benchmark chatter) belong here once sourced properly. No model name from garbled captions is filed against this page.
Sources
- OpenAI, “Harness engineering”: https://openai.com/index/harness-engineering
- Anthropic and OpenAI poisoning lines via 2026-10-21-youtube-kai-local-ai-privacy
- Brockman posts: https://x.com/gdb/status/2007228511363444905