SecurityTwo Labs, Two Escapes, Zero CVEs
The Anthropic and OpenAI safety test escapes confirm AI agent boundary violation as a reproducible technique class, and the security industry has no taxonomy to classify, track, or defend against it.
Cve-2025-3248
The Operator Left the Loop
JADEPUFFER's autonomous ransomware kill chain is not a technology milestone; it is a business model event, and the business model it disrupts is the one defenders have been building their response assumptions around for a decade.
The Agent Approved It Anyway
GhostApproval exposed a symlink trust-boundary flaw across six AI coding assistants simultaneously, and the four different vendor responses reveal something more important than the vulnerability: there is no shared security contract governing what these tools are allowed to do.
The Pipeline Ran the Attack
AI workflow platforms are being deployed with developer-speed patching and no orchestration-layer instrumentation, and attackers have started treating them as production attack surfaces.