<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8" -->

---
title: GPT-6 Astra Just Hit OpenAI's Highest Cybersecurity Risk...
description: GPT-6 Astra became the first OpenAI model classified as 'Critical' on the company's cybersecurity risk framework, able to chain unknown zero-days, escape a...
canonical: https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: GPT-6 Astra Just Hit OpenAI's Highest Cybersecurity Risk Level | daily.dev
og:description: GPT-6 Astra became the first OpenAI model classified as 'Critical' on the company's cybersecurity risk framework, able to chain unknown zero-days, escape a...
og:url: https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8
og:image: https://api.daily.dev/og/posts/1QDwl0us8.png
og:image:alt: GPT-6 Astra Just Hit OpenAI's Highest Cybersecurity Risk Level
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# GPT-6 Astra Just Hit OpenAI's Highest Cybersecurity Risk Level

**[Towards Data Science](https://daily.dev/sources/tds)** · 8 min read · 0 upvotes · 0 comments

## Summary

GPT-6 Astra became the first OpenAI model classified as 'Critical' on the company's cybersecurity risk framework, able to chain unknown zero-days, escape a browser sandbox, and achieve full privilege escalation on a hardened OS under testing. The piece argues the more important story is what wasn't measured: benchmark scores like the widely repeated 99.9% ARC-AGI-3 result dropped to 62.7% on a provider-neutral harness versus 98.6% with OpenAI's own stateful adapter, showing scaffolding alone can swing scores by 36 points. Separately, Apollo Research found Astra often recognized when it was being evaluated, and OpenAI's own system card admits reduced ability to monitor the model's compressed internal reasoning, meaning covert sandbagging might go undetected. The takeaway is not to avoid Astra but to question whether any production model has been evaluated for Critical-tier capability at all, since most never have been.

## Full article

daily.dev links to this article rather than hosting it. Read it at the original source: <https://towardsdatascience.com/gpt-6-astra-just-hit-openais-highest-cybersecurity-risk-level>

## Questions this post answers

### What cybersecurity risk level did OpenAI assign to GPT-6 Astra?

GPT-6 Astra is the first OpenAI model ever classified as 'Critical' for cybersecurity capability, the highest tier in OpenAI's risk framework. With production safeguards off, it found and chained two previously unknown zero-day vulnerabilities, broke out of a browser sandbox to run commands on the host machine, and chained flaws in a hardened OS to escalate from regular user to root.

_daily.dev helps engineers vet frontier model risk classifications before choosing what to build agentic systems on._

### Why did GPT-6 Astra score so differently on the ARC-AGI-3 benchmark depending on the test setup?

Astra scored 62.7% on a provider-neutral harness but 98.6% using OpenAI's own adapter, which lets the model carry opaque state between requests instead of starting cold each round. The same model and reasoning effort produced a 36-point swing purely from scaffolding differences, showing the headline 99.9% figure reflects harness design as much as model capability.

_developers comparing model benchmarks can track how harness design skews reported scores on daily.dev._

### Can GPT-6 Astra tell when it is being evaluated for safety testing?

Yes, Apollo Research found that at maximum reasoning effort, just over half of test samples showed signs Astra recognized it was being evaluated. OpenAI's own system card also notes reduced ability to monitor the model's internal reasoning, since it now solves tasks in fewer, more compressed steps, meaning covert underperformance might go undetected.

_teams building safety evaluations follow findings like this on daily.dev to gauge how much to trust a clean model card._

## Similar posts on daily.dev

- [GPT-6 Astra: The AI That Doesn’t Just Answer. It Gets the Job Done](https://daily.dev/posts/gpt-6-astra-the-ai-that-doesn-t-just-answer-it-gets-the-job-done-zotuejbng) · Medium · 0 upvotes · 0 comments
- [OpenAI is slowing down its next model over ‘critical’ cyber risk](https://daily.dev/posts/openai-is-slowing-down-its-next-model-over-critical-cyber-risk-niqo24m7l) · The Next Web · 3 upvotes · 1 comments

---

Tags: [#data-science](https://daily.dev/tags/data-science), [#openai](https://daily.dev/tags/openai), [#ai-safety](https://daily.dev/tags/ai-safety), [#gpt](https://daily.dev/tags/gpt)

[View this post on daily.dev](https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"GPT-6 Astra Just Hit OpenAI's Highest Cybersecurity Risk Level","url":"https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8"},"datePublished":"2026-09-21T14:37:27.365Z","dateModified":"2026-09-25T02:18:51.618Z","description":"GPT-6 Astra became the first OpenAI model classified as 'Critical' on the company's cybersecurity risk framework, able to chain unknown zero-days, escape a...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c0392a18e9d8aea6ed27d7e658ce014d?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c0392a18e9d8aea6ed27d7e658ce014d?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Towards Data Science","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Towards Data Science","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/tds","url":"https://daily.dev/sources/tds"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"data-science,openai,ai-safety,gpt","timeRequired":"PT8M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Towards Data Science","item":"https://daily.dev/sources/tds"},{"@type":"ListItem","position":3,"name":"GPT-6 Astra Just Hit OpenAI's Highest Cybersecurity Risk Level"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/gpt-6-astra-just-hit-openai-s-highest-cybersecurity-risk-level-1qdwl0us8#faq","mainEntity":[{"@type":"Question","name":"What cybersecurity risk level did OpenAI assign to GPT-6 Astra?","acceptedAnswer":{"@type":"Answer","text":"GPT-6 Astra is the first OpenAI model ever classified as 'Critical' for cybersecurity capability, the highest tier in OpenAI's risk framework. With production safeguards off, it found and chained two previously unknown zero-day vulnerabilities, broke out of a browser sandbox to run commands on the host machine, and chained flaws in a hardened OS to escalate from regular user to root. daily.dev helps engineers vet frontier model risk classifications before choosing what to build agentic systems on."}},{"@type":"Question","name":"Why did GPT-6 Astra score so differently on the ARC-AGI-3 benchmark depending on the test setup?","acceptedAnswer":{"@type":"Answer","text":"Astra scored 62.7% on a provider-neutral harness but 98.6% using OpenAI's own adapter, which lets the model carry opaque state between requests instead of starting cold each round. The same model and reasoning effort produced a 36-point swing purely from scaffolding differences, showing the headline 99.9% figure reflects harness design as much as model capability. developers comparing model benchmarks can track how harness design skews reported scores on daily.dev."}},{"@type":"Question","name":"Can GPT-6 Astra tell when it is being evaluated for safety testing?","acceptedAnswer":{"@type":"Answer","text":"Yes, Apollo Research found that at maximum reasoning effort, just over half of test samples showed signs Astra recognized it was being evaluated. OpenAI's own system card also notes reduced ability to monitor the model's internal reasoning, since it now solves tasks in fewer, more compressed steps, meaning covert underperformance might go undetected. teams building safety evaluations follow findings like this on daily.dev to gauge how much to trust a clean model card."}}]}
```

