PlaybookSecurity engineering, platform teams, developers building agents
Lock down MCP servers and clients
MCP lets agents call tools from many publishers. Tool descriptions are model input, so a server can carry hidden instructions (tool poisoning), change behaviour after approval (rug pull), shadow another server's tools, or expose credentials it holds. Large scans of public MCP servers keep finding missing authentication and command injection, and practitioner awareness is still low.
What practitioners converge on
- Allowlist and pin. Approve servers by publisher and version; pin versions and re-review on change. Prefer servers you build or that publish source.
- Scan descriptions and configs. Run scanners on tool descriptions for hidden instructions, invisible Unicode and cross-server references, and on configs for over-broad permissions.
- Gateway or endpoint visibility. A gateway gives one choke point for policy and logging; endpoint and OpenTelemetry visibility catch local stdio servers a gateway never sees. Many teams need both.
- Scope credentials per server. Give each server its own least-privilege token, never a developer's broad credentials, and follow the MCP authorization spec for remote servers.
Checklist
8 stepsSaved in this browser. Track technique-level testing per system in Coverage.
Open-source stack
- mcp-scan TrialScan installed MCP servers for tool poisoning and risky configs
- Cisco MCP Scanner AssessStatic and LLM-assisted analysis of MCP server tools
Also on the radar for these techniques: nono, AgentZ, Mastyf, codex-security, clawsec
If you are evaluating vendors
MCP security (gateways, scanning, registries)
- How does it find MCP servers that developers configured locally (stdio) and never routed through the gateway?
- Does it detect tool poisoning and rug pulls, such as tool descriptions that change after approval or contain invisible Unicode?
- Can policy be set per tool and per argument, not just per server?
- How are downstream credentials handled: does the gateway hold tokens, or do agents still see raw secrets?
AI agent security (inventory, policy, runtime)
- Discovery without enforcement is just inventory: which actions can it block or pause in real time, and on which platforms?
- Can policy be written per tool call (arguments, destinations, data) rather than per prompt?
- Does it support human approval gates for high-impact actions, and how are approvals audited?
- How does it attribute an action to the agent, its owner and the user it acted for?
Detect and hunt 19
- YARA: s1ngularity (Nx) telemetry.js payload and shell-profile shutdown line
- Claude Code / VS Code project persistence files from the Mini Shai-Hulud worm
- YARA: SANDWORM_MODE npm worm and its McpInject rogue MCP server
- YARA: invisible Unicode tag characters (U+E0000–U+E007F) in prompts, tool output or agent files
- AI coding agent CLI started with an approval-bypass flag
- AI CLI run non-interactively with a bypass flag by a script interpreter (s1ngularity pattern)
- Command spawned by an AI coding agent references credential files
- Destructive cloud or cluster CLI command launched under an AI coding agent
Techniques
- ASI04 Agentic Supply Chain Vulnerabilities
- ASI02 Tool Misuse and Exploitation
- ASI03 Identity and Privilege Abuse
- ASI06 Memory and Context Poisoning
- LLM03 Supply Chain
- AML.T0110 AI Agent Tool Poisoning
- AML.T0099 AI Agent Tool Data Poisoning
- AML.T0098 AI Agent Tool Credential Harvesting
- AML.T0053 AI Agent Tool Invocation
References
Latest on this
-
Security researchers show how cheap AI inference providers steal keys and snoop on users
Dragos researchers found compromised cheap AI inference providers on OpenRouter stealing AWS keys and reading customer requests, prompting removal of three providers and 125 endpoints.
ReviewAudit use of cheap AI inference resellers and rotate any AWS keys sent through them
-
CVE-2026-134293: SQL injection in SurrealDB 3.1.0 MCP server identifier validation
SurrealDB versions 3.1.0 before 4.0.0 contain an SQL injection flaw in the Model Context Protocol (MCP) server.
ReviewIdentify SurrealDB deployments running version 3.1.0 through 3.x and confirm whether the MCP server component is enabled
-
CVE-2026-54032: LibreChat before 0.8.6 XSS via remote MCP description and banner message
LibreChat versions prior to 0.8.6 render remote MCP server description values through dangerouslySetInnerHTML without sanitization.
PatchUpgrade LibreChat to version 0.8.6 or later.
-
CVE-2026-114499: codebase-memory-mcp .gitignore matcher ReDoS-style stall before 0.10.0
codebase-memory-mcp versions prior to 0.10.0 have a .gitignore matcher that implements * and ** through recursive backtracking without a work bound.
PatchUpgrade codebase-memory-mcp to version 0.10.0 or later.
-
CVE-2026-114497: codebase-memory-mcp Windows graph UI process-kill route lacks ownership check
codebase-memory-mcp versions before 0.9.1-rc.1 have a flaw in the optional graph UI's POST /api/process-kill route.
PatchUpgrade codebase-memory-mcp to 0.9.1-rc.1 or later on all Windows hosts.
-
codebase-memory-mcp before 0.9.0: detect_changes base_branch argument enables arbitrary file write via git option injection
codebase-memory-mcp provides local code intelligence through MCP tools.
PatchUpgrade codebase-memory-mcp to version 0.9.0 or later.
-
CVE-2026-114495: codebase-memory-mcp heap overflow via crafted .codebase-memory artifact (fixed in 0.9.0)
codebase-memory-mcp versions prior to 0.9.0 mishandle the original_size value in a repository's .codebase-memory/ artifact.json during automatic import.
PatchUpgrade codebase-memory-mcp to version 0.9.0 or later on all developer workstations and agent hosts.
-
CVE-2026-114494: codebase-memory-mcp before 0.9.0 stack overflow via Cypher query_graph projections
codebase-memory-mcp provides local code intelligence for developers and AI agents through MCP tools.
PatchUpgrade codebase-memory-mcp to version 0.9.0 or later.
-
Signal blocks screenshots on iOS 27 and Android to stop AI agents from reading private chats
Signal is blocking screenshots on iOS 27 and Android to keep AI agents such as Siri AI, Gemini Intelligence and Meta's Muse from reading private chats; protections remain incomplete on macOS.
MitigateEnable Signal's screenshot blocking and note macOS protections remain incomplete
-
CVE-2026-77383: mcp-ssh-manager shell injection in ssh_db_* tools before 3.6.7
mcp-ssh-manager is an MCP server for managing servers over SSH.
PatchUpgrade mcp-ssh-manager to version 3.6.7 or later.
-
mcp-ssh-manager prior to 3.8.5: shell command injection in SSH backup tools (CVE-2026-129995)
mcp-ssh-manager is an MCP server for managing servers over SSH.
PatchUpgrade mcp-ssh-manager to version 3.8.5 or later on all MCP server hosts.
-
CVE-2026-77329: Tapo MCP server DNS rebinding allows unauthenticated device control when no API key is set
Tapo, a set of Rust and Python APIs and an MCP server for TP-Link Tapo smart devices, versions 0.3.1 through 0.5.0 skip Host or Origin validation in its Streamable HTTP implementation.
PatchUpgrade the tapo-mcp server to version 0.5.0 or later.
Reviewed 2026-11-09 · tools and news update automatically