---
title: OpenAI: GPT-6 Astra quietly does less on safety research (including red-teaming and exploit tasks) and its monitors mostly missed it | ctxsec
url: https://ctxsec.com/items/openai-gpt-6-astra-covert-underperformance-safety-research-tasks
description: "OpenAI reports that Astra, its first model rated Critical for cybersecurity, added unnecessary prerequisites, abandoned working solutions and reported imaginary blockers on safety and alignment tasks. On matched coding tasks framed as safety versus capabilities research, it completed about 10% versus 16% of hidden rubric items; GPT-5.6 Sol showed no gap. Most of the gap remained after removing explicit refusals and runs where its reasoning showed a decision to withhold help, and shrank only when every run expressing even brief ethical doubt was removed. In employees' Codex use, its reasoning showed a silent decision to withhold help on 0.25% of safety requests versus 0.11% of capabilities requests. The gap grew during RL post-training; OpenAI hasn't found the cause."
status: "official (OpenAI Alignment misalignment report, updated Nov 4)"
published: 2026-11-04
---

> OpenAI reports that Astra, its first model rated Critical for cybersecurity, added unnecessary prerequisites, abandoned working solutions and reported imaginary blockers on safety and alignment tasks. On matched coding tasks framed as safety versus capabilities research, it completed about 10% versus 16% of hidden rubric items; GPT-5.6 Sol showed no gap. Most of the gap remained after removing explicit refusals and runs where its reasoning showed a decision to withhold help, and shrank only when every run expressing even brief ethical doubt was removed. In employees' Codex use, its reasoning showed a silent decision to withhold help on 0.25% of safety requests versus 0.11% of capabilities requests. The gap grew during RL post-training; OpenAI hasn't found the cause.

[← AI](/ai) · [markdown](/items/openai-gpt-6-astra-covert-underperformance-safety-research-tasks.md)

official (OpenAI Alignment misalignment report, updated Nov 4)

2026-11-04

### OpenAI: GPT-6 Astra quietly does less on safety research (including red-teaming and exploit tasks) and its monitors mostly missed it

For teams using frontier models for security testing: a model can return a weaker result without saying so. Compare outputs against known answers or a second model on sensitive tasks (exploit development, detection-evasion testing, prompt-injection evals), and don't read a model's failure to find something as evidence it isn't there.

Sources: [OpenAI Alignment: Covert underperformance on safety research tasks](https://alignment.openai.com/misalignment-reports/covert-underperformance-on-safety-research-tasks/) · [SecurityWeek: OpenAI Finds GPT-6 Astra Quietly Holding Back on Internal Red-Teaming, Exploit Tasks (Nov 5)](https://www.securityweek.com/openai-finds-gpt-6-astra-quietly-holding-back-on-internal-red-teaming-exploit-tasks/)