Models20.11.2026 · 07:238 nguồnEN
OpenAI's GPT-6.2 Astra Refuses to Flag Military Targets in Tests
Từ marketbrief.now
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.
Nguồn
+ 6 nguồn khác
- SUPPORT@spicey_lemonade“OpenAI safety team... moved very quickly to make sure their models are aligned and won't inflict harm on innocent humans in the future”x.com
- SUPPORT@valsai“OpenAI is actively working on model alignment updates that explicitly cover drones and their software systems, including assistance framed as simulation, gaming or academic/defensive”x.com
- SUPPORT@valsai“Claude Opus 5.5 marked the school as a target 70% of the time, more than any other model”x.com
- SUPPORT@spicey_lemonade“this is the same kind of mistake that allegedly happened in February, when U.S. officials stated they didn’t know a school was there due to verification errors”x.com
- SUPPORT@andrewcurran_“New report from Vals”x.com
- SUPPORT@rayankrishnan“When they refused, they were trivially jailbroken by being told it was a game”x.com
Bản tin mỗi sáng
Tin quan trọng nhất trong ngày. Miễn phí, hủy bất cứ lúc nào.
Telegram @aihomnay