High-signal AI/security/automation notes.
MCPTox benchmark evaluates 20 LLM agents against tool poisoning attacks on 45 real MCP servers — o1-mini hit 72.8% success rate, with more capable models often more vulnerable.
The TeamPCP supply-chain operation expanded from SAP npm packages to PyPI (PyTorch Lightning) and intercom-client within 48 hours, targeting AI and ML developer ecosystems.
Novee researchers discovered a high-severity vulnerability in Cursor IDE (CVE-2026-26268, CVSS 8.1) where the AI coding agent autonomously triggers malicious Git hooks when opening attacker-controlled repositories.
NSFOCUS announced AI-Scan, a dedicated security scanner for the OpenClaw agent ecosystem covering CVE matching, credential leak detection, memory poisoning, and supply-chain risk.
New research proves embedding-based RAG defenses miss 79-90% of numerical manipulations — changing a tax figure by $50,000 yields 0.9998 cosine similarity, invisible to detection.
Wiz launches Red Agent for AI-driven vulnerability simulation, AI-BOM for framework inventory, and Wiz Code guardrails — expanding its AI Application Protection Platform from code to cloud edge.
Most AI security scanners only cover packaged dependencies, missing the dynamic components agents load at runtime — model adapters, MCP connections, and first-invocation tools.
A new benchmark tests five frontier LLMs on real-world threat hunting across 106 attack procedures — the best model flags only 3.8% of malicious events.
Four MCP trust boundaries — permissions, tools, inputs, and observability — that teams are getting wrong when connecting AI agents to real systems.
Critical RCE in Pipecat, the open-source Python framework for voice and multimodal AI agents, via insecure pickle deserialization in LivekitFrameSerializer.
New arXiv paper introduces T-MAP, a red-teaming framework that achieves 57.8% attack success rate against frontier LLM agents through trajectory-aware evolutionary search in MCP environments.
Cisco researchers demonstrated a memory poisoning attack against Claude Code that achieves persistent behavioral manipulation across all projects and sessions via npm supply-chain hooks.
Critical pre-auth SQL injection in LiteLLM (CVE-2026-42208) was exploited within 36 hours of disclosure, enabling attackers to extract virtual API keys and provider credentials from the PostgreSQL backend.