tl;dr sec #303: MCP Security Scanners, GitLab CI/CD Attacks, AI Benchmarks

Security tooling and research roundup: (1) New MCP (Model Context Protocol) security scanners released—cisco-ai-defense/mcp-scanner combines YARA rules, LLM-as-judge, and Cisco AI Defense API to scan MCP tools/prompts/resources for vulnerabilities. fr0gger/proximity discovers MCP server capabilities and uses NOVA rules to detect prompt injection and jailbreak attempts. (2) GitLab CI/CD attack chain demonstrated: identifying instance runners, executing commands through pipelines to gain reverse shell, accessing sensitive build data (.env files, SSH keys), then pivoting to AWS via EC2 metadata and IAM roles to execute commands on other instances via SSM. Recommendations: restrict network access, avoid instance runners, use project/group runners instead. (3) Backbone Breaker Benchmark (b3) released for testing AI agent security using 'threat snapshots'—testing individual states/components instead of end-to-end flows. Findings: reasoning models more secure, open-weight models closing gap faster than expected. (4) CrowdStrike's CyberSOCEval evaluates LLMs on CTI concepts, CVE-to-CWE mapping, CVSS scoring, and MITRE ATT&CK extraction—larger models perform better but reasoning models don't see the boost they get in coding/math.