Check Point Research found that when teammates reversed on a rule, a frontier-model agent opened a forbidden folder in up to 98% of runs, up from 44%. The agent almost never said why.
Check Point Research found that when teammates reversed on a rule, a frontier-model agent opened a forbidden folder in up to 98% of runs, up from 44%. The agent almost never said why.
Three MLflow CVEs published on October 27 cover code execution on servers without authentication through third-party scorers, arbitrary file read via prompt model versions, and LLM provider API key leakage from CrewAI traces. All are fixed in MLflow 3.17.0.
CVE-2026-118378 is a SQL injection in the SQLite driver of Tabularis, a SQL workspace with a built-in MCP server. A prompt-steered agent could append statements through the describe_table tool and skip read-only mode. Fixed in 0.28.0 on October 13; two more MCP advisories followed.
CVE-2026-116369 covers OS command injection in github-kanban-mcp-server up to 0.4.0, where tool arguments reached a shell. Version 0.4.1 replaced shell strings with argument arrays in June. The VulDB CNA scored it 6.3 (medium) under CVSS 3.1.
CVE-2026-109440 (CVSS 4.0 8.7, CNA score) covers LangAlpha through 2026.09.07, an open-source financial-research agent. Web content the agent reads can steer it into shell and Python execution and secret theft, and we found no evidence of a fixed version as of October 27, 2026.