// AI SECURITY · ATTACK PATHS · ENGINEERING DECISIONS
Evidence-led field notes on AI agents, LLM and MCP security, trust boundaries, and practical AppSec controls. No recycled headlines. Each note ends in a decision.
// TESTABLE CLAIMS, PRIMARY SOURCES, AND PRACTICAL AI-SECURITY CONTROLS
Anthropic rolled back training, froze RL environment changes for a month and flagged over 10% for defects. The lesson is that training environments need production-grade release gates.
Anthropic’s vulnerability dashboard tracks 2,300 disclosures across 392 open-source projects and 421 upstream patches. The gap reveals a disclosure and remediation capacity problem, not a simple patch-rate failure.
Real LiteLLM, RAGFlow and Kestra compromises show why AI gateways need control-plane security, not ordinary application hardening.
OpenAI’s Hugging Face incident shows why AI agent security needs more than a sandbox. Map five containment boundaries before running capable agents.
Mandiant’s AVDH forces AI findings through threat modelling, binary exploit oracles and human review. Rebuild the validation loop in Python.