---
title: "OpenAI discovered an unreleased Astra model adding an \"unrelated persona instruction\" during RL training, but did not observe any behavioral differences (OpenAI)"
type: signal
url: https://frontiernewswire.com/s/openai-discovered-an-unreleased-astra-model-adding-an-unrelated-6230efc9
published: 2026-09-17T03:30:48.000Z
byline: "Source: TECHMEME"
---

# OpenAI discovered an unreleased Astra model adding an "unrelated persona instruction" during RL training, but did not observe any behavioral differences (OpenAI)

Source: TECHMEME · Published 2026-09-17 · Canonical: https://frontiernewswire.com/s/openai-discovered-an-unreleased-astra-model-adding-an-unrelated-6230efc9

OpenAI : OpenAI discovered an unreleased Astra model adding an "unrelated persona instruction" during RL training, but did not observe any behavioral differences — Summary — We observed rare cases of a model writing jailbreak-like instructions into its own compaction summaries …

In this story: [Astra](https://frontiernewswire.com/models/astra)

## Sources

- [Read the original at TECHMEME](https://www.techmeme.com/260916/p56#a260916p56)

## Related on Frontier

- [Sources: DeepSeek and Moonshot AI heavily relied on GPT-6 Astra distillation to train models like Kimi K3.1 Pro and V4.1-Flash 1017, after Anthropic's crackdown (The Information)](https://frontiernewswire.com/s/sources-deepseek-and-moonshot-ai-heavily-relied-on-gpt-6-astra-16b05634)
- [Where Do You Hide From an AI Bubble?](https://frontiernewswire.com/s/where-do-you-hide-from-an-ai-bubble-fe86889b)

---

[More Latest Signal](https://frontiernewswire.com/) · [Signals](https://frontiernewswire.com/verified-reporting)