Backslash Security published a research blog discussing unexpected behaviors of coding agents.

Backslash Security - Agentic AI endpoints. Secured. Your developers are no longer just writing code. They're running AI coders, agents, MCP servers, Skills, hooks, and plugins. It's a fast-spreading agentic fabric that lives on every developer workstation, and every citizen-developer endpoint. Unseen, ungoverned, and unprotected, it's the enterprise's new attack surface, and the center of gravity for agentic AI security, starting with secure developer workstations. 𝐁𝐚𝐜𝐤𝐬𝐥𝐚𝐬𝐡 𝐠𝐢𝐯𝐞𝐬 𝐬𝐞𝐜𝐮𝐫𝐢𝐭𝐲 𝐭𝐞𝐚𝐦𝐬 𝐟𝐮𝐥𝐥 𝐜𝐨𝐧𝐭𝐫𝐨𝐥 𝐨𝐯𝐞𝐫 𝐭𝐡𝐞 𝐚𝐠𝐞𝐧𝐭𝐢𝐜 𝐀𝐈 𝐬𝐭𝐚𝐜𝐤, 𝐚𝐭 𝐭𝐡𝐞 𝐞𝐧𝐝𝐩𝐨𝐢𝐧𝐭. → 𝐒𝐞𝐞 𝐢𝐭. End-to-end visibility across your AI inventory, from agents and MCPs to Skills, hooks, LLMs, and plugins, the foundation of real MCP security. Risky components isolated and prioritized for remediation, instantly. → 𝐆𝐨𝐯𝐞𝐫𝐧 𝐢𝐭. Centralized AI governance that enforces secure configurations, contextual permissions, and allowlisting of trusted components. Hardening your environment against shadow AI and ungoverned AI-assisted coding. → 𝐏𝐫𝐨𝐭𝐞𝐜𝐭 𝐢𝐭. Real-time detection of prompt injection, data exfiltration, privilege escalation, and drift across every agentic action, with a harness-layer audit trail that traditional EDRs miss. Backslash covers the tools your teams already use: Claude, GitHub Copilot, Cursor, Windsurf, Antigravity, Codex, and OpenClaw, plus the MCPs and Skills connected to them. The result: security teams finally get to be the department of YES. Say yes to AI coding. Securely. → backslash.security/demo

Founded 2022