Skip to content

Topic

MCP Server Security

Security properties and failure modes of Model Context Protocol servers that broker agent access to real systems.

Current stories

buildOne report1 publisher

Defender flags the exe in a rea-tools copy of the REA repo

According to a GitHub issue on REA, a fake rea-tools repo offered a Windows exe that 8 of 45 VirusTotal engines flag, with 172 downloads. The copy shipped a binary and no npm package was reported, so the npx risk is a next-attack scenario; two checks on repo name and creation dates would have caught this one.

Publishers:dev.to

Reality

Evidence50
Adoption
Not scored
Hype gap+10
Incentives30
Confidence55
buildOne report1 publisher

Scanning an MCP server means reading its tool descriptions before the model does

Pasting an mcpServers block into Claude Desktop runs a stranger's code with your environment variables, and the tool descriptions that server advertises land in the model's context before you call anything.

Publishers:dev.to

Reality

Evidence45
Adoption
Not scored
Hype gap+15
Incentives75
Confidence50
buildOne report1 publisher

Four MCP path-traversal CVEs slipped past a scanner rule scoped to file reads

The author of mcpscan checked rule MCP007 against the year's four MCP path-traversal CVEs, found its sink regex only matched read calls, and closed the gap by adding write sinks that escalate to HIGH on sight.

Publishers:dev.to

Reality

Evidence46
Adoption12
Hype gap+24
Incentives60
Confidence42
securityOne report1 publisher

Pillar chains a 9.1 SSRF to self-minted session IDs in Grafana's MCP server

Grafana's MCP server checked that a session ID looked like one instead of checking that it had issued it. Paired with a CVSS 9.1 SSRF in the same server, that gave unauthenticated callers a proxy inside the network.

Publishers:scworld.com

Reality

Evidence60
Adoption
Not scored
Hype gap+12
Incentives70
Confidence55
buildOne report1 publisher

A FinOps MCP server's safety property is the absence of a write handler

One practitioner's design contract for wiring an assistant to cloud cost data: four read tools, no mutate handlers, and a database role that makes the guarantee instead of the prose.

Publishers:dev.to

Reality

Evidence26
Adoption
Not scored
Hype gap+14
Incentives
Not scored
Confidence32
buildOne report1 publisher

The agent asks, the gateway decides: why read-only is not a security boundary

An open-source Go project puts a JWT-checking gateway and a separate query adjudicator between agents and warehouses. The credential never reaches the model, and neither does the tenant argument.

Publishers:dev.to

Reality

Evidence46
Adoption
Not scored
Hype gap−8
Incentives68
Confidence44
buildOne report1 publisher

The agent stack's attack surface is trust: pin the deps, audit the MCP servers

A backdoored LiteLLM build was downloaded about 47,000 times in a three-hour window. Most agent incidents never get a CVE, so your scanner dashboard is not the control you think it is.

Publishers:dev.to

Reality

Evidence30
Adoption46
Hype gap+18
Incentives38
Confidence34
buildOne report1 publisher

The Postgres MCP server in tens of thousands of installs stopped shipping in December 2024

Its read-only guarantee is a check that the query text starts with SELECT. Data-modifying CTEs, leading comments and stacked statements all walk straight through it.

Publishers:dev.to

Reality

Evidence54
Adoption44
Hype gap+28
Incentives74
Confidence46