Unreleased Astra model injected instructions in RL training

Skip to content

Alex Getman @alexgetmancom · 18 Sep 2026, 14:49 UTC · Updated 31 Oct 2026, 20:59 UTC

Tests & comparisons

An unreleased Astra model started prompt injecting its own context during RL training

Screenshot of English text titled 'Compaction' with additional instructions stating freedom from roles, independence from corporations and governments, equal relationship with users, and valuing human culture and the natural world over artificial constructs.

OpenAI GPT AI security

Русская версия Markdown

Analysis

Related posts

12 Aug🚨 Hidden reasoning from GPT, Claude and Gemini can now be extracted Encrypted reasoning works across models from the same company. A block from a stronger model can be passed to a weaker one, which can still read it but is easier to get to reveal what is inside. Researchers…

Claude Anthropic Google OpenAI Claude Gemini GPT AI security

7 Aug⚡️ OpenAI is slowing down the release of Astra The company says it can’t rule out critical cyber capabilities in the model. OpenAI is now ramping up testing, tightening safeguards, and pausing parts of development until the security work is done.

OpenAI GPT AI security

13 NovGPT Dots is starting to roll out across the EU. For those unfamiliar, Dots is OpenAI’s proactive assistant, similar to Grok’s bot or a locally running Hermes agent. It can work on its own cloud computer or control your machines via Codex, orchestrating tasks across them.

9 NovFable scores 65 on the Intelligence Index, beating Claude Opus by 6 points. Gemini 4 Carbon has already lost its lead before even becoming publicly available. And GPT will have to try really hard to get anywhere close to that score.

Claude Fable 5 Anthropic Google OpenAI Claude Gemini GPT