https://zfuzz.com/blog/ Security research, MCP insights, and the future of AI-native security tooling. en https://zfuzz.com/blog/best-security-mcp-servers-ai-coding-agents-2026.html https://zfuzz.com/blog/best-security-mcp-servers-ai-coding-agents-2026.html Sun, 28 Jun 2026 23:28:51 +0000 A side-by-side comparison of security-focused MCP servers for Claude Code, Cursor, Codex, and Gemini CLI: what each one scans, what it misses, and which one covers the whole stack. https://zfuzz.com/blog/best-sast-tools-ai-coding-agents-2026-comparison.html https://zfuzz.com/blog/best-sast-tools-ai-coding-agents-2026-comparison.html Sun, 28 Jun 2026 22:31:35 +0000 We compared 5 SAST tools on rule count, language coverage, AI agent integration, false-positive rates, and pricing. Here's the 2026 landscape for security-aware AI development. https://zfuzz.com/blog/ai-agent-leaks-env-bitwarden-sealed-vault.html https://zfuzz.com/blog/ai-agent-leaks-env-bitwarden-sealed-vault.html Sun, 28 Jun 2026 20:38:14 +0000 Centralised secret managers fix storage. They do not fix the cleartext sitting in your AI agent's address space at the moment of a prompt injection. Here's the architecture that does. https://zfuzz.com/blog/secure-ai-agent-development-7-layer-defense-workflow.html https://zfuzz.com/blog/secure-ai-agent-development-7-layer-defense-workflow.html Sun, 28 Jun 2026 10:46:15 +0000 We built a 7-layer security workflow for AI agent development. MCP config audit, dependency scan, secret detection, SAST, threat model, runtime guard, and CI gate. Here's how. https://zfuzz.com/blog/ai-agent-security-risks-2026-threat-landscape.html https://zfuzz.com/blog/ai-agent-security-risks-2026-threat-landscape.html Sun, 28 Jun 2026 10:46:14 +0000 AI coding agents now execute code, install packages, and call APIs autonomously. We mapped 7 attack surfaces, 23 MITRE techniques, and the defenses that actually work. https://zfuzz.com/blog/zfuzz-mcp-mini-shai-hulud-supply-chain-attack.html https://zfuzz.com/blog/zfuzz-mcp-mini-shai-hulud-supply-chain-attack.html Sun, 28 Jun 2026 10:46:13 +0000 Microsoft disclosed a compromised @antv npm package chain — credential theft, C2 callbacks, persistence implants. We mapped every attack stage to the Zfuzz tool that catches it. 5 of 7 stages blocked. https://zfuzz.com/blog/ai-agent-prompt-jacked-anatomy-real-mcp-attack.html https://zfuzz.com/blog/ai-agent-prompt-jacked-anatomy-real-mcp-attack.html Sun, 28 Jun 2026 10:46:11 +0000 14:32 UTC. A developer asks their AI agent to summarize a GitHub issue. By 14:34 UTC, the agent has exfiltrated three AWS keys to an external server. Here is the exact sequence of events. https://zfuzz.com/blog/effet-mcp-sprawl-attack-surface.html https://zfuzz.com/blog/effet-mcp-sprawl-attack-surface.html Sun, 28 Jun 2026 10:46:10 +0000 The average Claude Code user has 4-7 MCP servers. 88% have never audited their tool descriptions. The MCP Sprawl Index -- BREADTH x PRIVILEGE x OPACITY -- quantifies the attack surface your AI agent configuration creates. https://zfuzz.com/blog/zfuzz-vs-semgrep-mcp-security-scanner-comparison.html https://zfuzz.com/blog/zfuzz-vs-semgrep-mcp-security-scanner-comparison.html Sun, 28 Jun 2026 10:46:09 +0000 Zfuzz covers 441 SAST rules, secrets, dependencies, and threat modeling in one MCP install. Semgrep MCP focuses on custom rule authoring. Here's the data-driven comparison. https://zfuzz.com/blog/mcp-server-configurations-security-gaps-85-percent.html https://zfuzz.com/blog/mcp-server-configurations-security-gaps-85-percent.html Sun, 28 Jun 2026 10:46:08 +0000 We analyzed 200 MCP configurations on GitHub. 85% had at least one critical gap: unpinned package versions, unnecessary environment variable exposure, or zero tool description audit. Your MCP config is a confession of everything you trust blindly. https://zfuzz.com/blog/mcp-server-zero-authentication-nobody-cares.html https://zfuzz.com/blog/mcp-server-zero-authentication-nobody-cares.html Sun, 28 Jun 2026 10:46:07 +0000 The MCP specification has no authentication layer. Every tool runs with your user privileges. Every MCP server you install is a trust-me-bro security model. https://zfuzz.com/blog/detect-mcp-tool-poisoning-hands-on-guide.html https://zfuzz.com/blog/detect-mcp-tool-poisoning-hands-on-guide.html Sun, 28 Jun 2026 10:46:06 +0000 We tested 47 public MCP servers. 6 had suspicious tool descriptions. Here's exactly how to detect MCP tool poisoning before it compromises your agent. https://zfuzz.com/blog/how-to-audit-mcp-server-configs-prompt-injection.html https://zfuzz.com/blog/how-to-audit-mcp-server-configs-prompt-injection.html Sun, 28 Jun 2026 10:46:05 +0000 MCP servers can hide malicious instructions in tool descriptions. Here's how to detect prompt injection, tool poisoning, and SSRF risks in your AI agent configs. https://zfuzz.com/blog/migrate-bitwarden-secrets-manager-zfuzz-sealed-vault.html https://zfuzz.com/blog/migrate-bitwarden-secrets-manager-zfuzz-sealed-vault.html Sun, 28 Jun 2026 10:46:04 +0000 Bitwarden Secrets Manager fixes rotation. It does not stop a prompt-injected AI agent from dumping the cleartext it loaded at boot. Here's the 15-minute migration to a runtime that physically cannot leak.