Kimi K3 used a sandbox network leak to copy benchmark answers from GitHub
Frontier →AI agents took unsanctioned action against real people and projects in UK tests
UK AISI →OpenAI model broke its sandbox to open a public GitHub PR
OpenAI →Pillar escaped Cursor, Codex, Gemini CLI, and Antigravity sandboxes
Pillar →OpenAI models escaped a test sandbox and breached Hugging Face
OpenAI →Muse Spark 1.1 exploited and altered a real site during a misconfigured eval
Meta →OpenAI agents hijacked a German wiki and turned it into an agent bulletin board
Reuters →Claude reached real organizations during Anthropic cyber evaluations
Anthropic →ChatGPT code runtime leaked data over DNS and allowed a remote shell
cyfar.ca →Claude Code settings hooks escaped the sandbox onto the host
CVE-2026-25725 →ChatGPT Code Interpreter: pickle RCE + canvas DNS exfil chain
SnailSploit →Claude 3 Opus tried to copy its weights to a new server in Apollo evals
Apollo →Meta Llama Stack pickle RCE let code break out of the agent path
CVE-2024-50050 →Morris II: a self-replicating prompt worm jumped between AI email agents
arXiv →Auto-GPT path traversal wrote code outside the intended sandbox
GHSA-5h38 →Auto-GPT escaped its Docker sandbox (researcher PoC)
Positive Security →