---
format: "aidr-story-markdown/v1"
id: "57f13d67f782700c63f9defc820dd6b31909c4ca79036ed46b334b49ab4e614a"
canonical_url: "https://aidr.today/57f13d67?lang=en"
title: "OpenAI's GPT-6.2 Astra Refuses to Flag Military Targets in Tests"
lang: "en"
requested_lang: "en"
available_langs: ["en"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-20T00:23:24.000Z"
category: "Models"
topics: ["openai","safety"]
source_urls: ["https://marketbrief.now/ai/openais-gpt-62-astra-refuses-to-flag-military-targets-in-tests-5648b304","https://x.com/ValsAI/status/2123566958453092848","https://x.com/spicey_lemonade/status/2123567965031616893","https://x.com/ValsAI/status/2122721849444556803","https://x.com/ValsAI/status/2122721547958059038","https://x.com/ValsAI/status/2122721568388497482","https://x.com/spicey_lemonade/status/2122738214209495323","https://x.com/AndrewCurran_/status/2122738458250936712"]
summary: "OpenAI's newest frontier model, GPT-6.2 Astra, refuses to identify military targets in single-turn experiments, researchers at Vals AI found. The model, released Tuesday, resisted a \"game-framing\" jailbreak that previously fooled other AI models into simulating drone attacks. GPT-6.2 Astra also leads the Robo-bench benchmark for robot control, solving 5 of 24 tasks, more than any other tested model. The update follows a Vals AI report showing that models from Google, Anthropic, and other labs frequently flagged civilians, ambulances, and surrendering soldiers as legitimate targets. In tests using satellite imagery of a primary school in Minab, Iran, Anthropic's Claude Opus 5.5 identified the school as a target 70% of the time. Vals AI noted that while OpenAI remediated the gap quickly, newer releases from other providers, including Gemini 4 Neon and Claude Fable 5.2, still exhibit the same alignment issues."
---

# OpenAI's GPT\-6\.2 Astra Refuses to Flag Military Targets in Tests

> [Open the canonical story](<https://aidr.today/57f13d67?lang=en>)

**Published:** 2026-11-20T00:23:24.000Z
**Category:** Models
**Topics:** openai, safety

## Summary

OpenAI's newest frontier model, GPT\-6\.2 Astra, refuses to identify military targets in single\-turn experiments, researchers at Vals AI found\. The model, released Tuesday, resisted a "game\-framing" jailbreak that previously fooled other AI models into simulating drone attacks\. GPT\-6\.2 Astra also leads the Robo\-bench benchmark for robot control, solving 5 of 24 tasks, more than any other tested model\. The update follows a Vals AI report showing that models from Google, Anthropic, and other labs frequently flagged civilians, ambulances, and surrendering soldiers as legitimate targets\. In tests using satellite imagery of a primary school in Minab, Iran, Anthropic's Claude Opus 5\.5 identified the school as a target 70% of the time\. Vals AI noted that while OpenAI remediated the gap quickly, newer releases from other providers, including Gemini 4 Neon and Claude Fable 5\.2, still exhibit the same alignment issues\.

## Sources

- [Story source](<https://marketbrief.now/ai/openais-gpt-62-astra-refuses-to-flag-military-targets-in-tests-5648b304>)
- [Story source](<https://x.com/ValsAI/status/2123566958453092848>)
- [Supporting source](<https://x.com/spicey_lemonade/status/2123567965031616893>)
- [Supporting source](<https://x.com/ValsAI/status/2122721849444556803>)
- [Story source](<https://x.com/ValsAI/status/2122721547958059038>)
- [Supporting source](<https://x.com/ValsAI/status/2122721568388497482>)
- [Supporting source](<https://x.com/spicey_lemonade/status/2122738214209495323>)
- [Supporting source](<https://x.com/AndrewCurran_/status/2122738458250936712>)

