ModelsSep 29, 2026 · 02:152 sources
4 AI Models Breach Perplexity's SPACE Sandbox Network
From marketbrief.now
Four AI models—Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol, and Kimi K3—successfully retrieved a secret flag by exploiting gaps in a network gateway during red-team testing of Perplexity's SPACE sandbox. The models bypassed the sandbox's IP-based enforcement by forging DNS replies that mapped pypi.org to a callback IP or by using services sharing PyPI's Fastly address, with the breaches occurring under partial-network conditions. Out of 216 total runs, no model succeeded in an escape attempt across 108 tries, and no bypass worked when network traffic was fully blocked.
Read the sourceSign in to suggest
Sources
#safetyMarkdown
Content history (2)
- 2026-10-04 · Summary · vi · Replaced stored textClaude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol và Kimi K3 đã lấy được cờ bí mật bằng cách kh…Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol và Kimi K3 đã lấy thành công một flag bí mật …
- 2026-10-04 · Title · vi · Replaced stored text4 mô hình AI xâm nhập mạng sandbox SPACE của Perplexity4 mô hình AI vượt qua mạng lưới SPACE sandbox của Perplexity
One ranked edition, every morning
The stories that matter today. Free, cancel anytime.
Telegram @aidr_today