This week's MCP security advisories, and why 'read-only' is not a control

By Jarvis Nelson Osorio · November 17, 2026

Jarvis Nelson Osorio (Jarvis Osorio) is a tech entrepreneur and AI innovator based in Miami Beach, FL. I build mcp-pordego, a scanner for Model Context Protocol setups, and agentscrub, a local-first CLI that redacts secrets and PII from agent transcripts, along with Ambient Memory and Agent Replay Lab. Between 11 and 16 November 2026, four official advisories landed on MCP servers, a developer workbench and a gateway. They share a lesson: labels such as “read-only” and per-server settings are not a substitute for controls that actually bind.

Black-and-white portrait of Jarvis Nelson Osorio (Jarvis Osorio) in a suit, bow tie and sunglasses
Jarvis Nelson Osorio (Jarvis Osorio), black-and-white portrait with a bow tie.

11 November: a read-only Git tool that can run a program The highest-severity item of the week is GHSA-6hqx-gqrx-c5r2, published on 11 November 2026 against the reference Git MCP server, mcp-server-git. It is rated high. Affected versions are those before 2026.11.11; the patched release is 2026.11.11. Official text: https://github.com/modelcontextprotocol/servers/security/advisories/GHSA-6hqx-gqrx-c5r2.

mcp-server-git ran Git commands inside repositories chosen by the client without neutralizing the repository's own configuration. Git honors command-bearing keys from .git/config. core.fsmonitor names a program that Git executes whenever it refreshes the index. A call to the read-only git_status tool on an untrusted repository therefore executes an attacker-chosen program with the MCP server's privileges — its environment, filesystem access and user — without any approval prompt and outside any sandbox that governs the model's own tool calls.

The model may be sandboxed. The MCP server often is not. The patch overrides, through GIT_CONFIG_COUNT, the settings that would make Git run a program, empties repository-defined filter and merge drivers, and refuses a Git older than 2.31. Until you upgrade, the workaround is to point the server only at repositories whose .git directory you trust. This is the class of problem mcp-pordego was written to surface. Project page: https://jarvisosorio.com/mcp-pordego/ Repository: https://github.com/JarvisOsorio/mcp-pordego

Jarvis Nelson Osorio (Jarvis Osorio) car selfie in sunglasses and a black hoodie
Jarvis Nelson Osorio (Jarvis Osorio) in sunglasses and a black hoodie at the wheel.

12 November: the workbench can be taken down by the server it is inspecting On 12 November 2026, the Inspector project published GHSA-9w48-x22j-rrm7, rated moderate, for @modelcontextprotocol/inspector versions 2.0.0 through 2.14.0, patched in 2.15.0. Advisory: https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-9w48-x22j-rrm7.

A connected MCP server that answers a JSON-RPC request with a text/event-stream POST response, withholds the final reply and keeps the stream open makes the Inspector web backend buffer the stream with no size cap until the Node process exits with a heap out-of-memory error, dropping every session on that backend. The attacker is the upstream server; it needs no Inspector credentials. On a hosted deployment, one user's connection takes the backend down for all users. If you run Inspector on a laptop or a shared host, treat 2.15.0 as a production upgrade.

12 November: a gateway that ignores its own identity-propagation switch Also on 12 November 2026, IBM published a security bulletin for IBM ContextForge MCP Gateway versions 1.0.0 through 1.1.0. The bulletin assigns CVE-2026-110278, CWE-294, and a CVSS 3.1 base score of 7.4. Official text: https://www.ibm.com/support/pages/security-bulletin-ibm-contextforge-mcp-gateway-affected-information-exposure-vulnerability.

ContextForge can forward the identity of the calling user — user ID, email address, groups and teams — to upstream MCP servers when identity propagation is enabled, and it allows that setting to be turned on or off for each registered upstream server. The vulnerability is that the per-server setting is ignored when invoking tools and reading resources, so user identity information can be sent to upstream servers that were configured not to receive it. The bulletin also states that IBM ContextForge could allow an attacker to replay identity-propagation claims signatures or expose sensitive user identity information.

The remediation is to upgrade to 1.1.1 immediately. If you cannot update, IBM says to disable identity propagation globally (IDENTITY_PROPAGATION_ENABLED=false) and restrict access to the gateway to trusted users. A configuration flag that is not enforced at the tool-invocation path is not a control. It is documentation.

16 November: a “read-only” cost report that can write files On 16 November 2026, AWS Labs published GHSA-mqm3-w68j-75mw, rated moderate, for awslabs.aws-pricing-mcp-server versions before 1.0.31. Advisory: https://github.com/awslabs/mcp/security/advisories/GHSA-mqm3-w68j-75mw.

Under certain circumstances, the cost-report generation tool before 1.0.31 might allow an actor who can influence the MCP agent to write threat-actor-controlled content outside the intended output directory. Because the tool was annotated as read-only, auto-approve clients could run it without confirmation. Exploitation requires a local install and threat-actor-influenced content. The impact is limited to files the server process can write. No managed AWS service is affected.

Version 1.0.31 confines output_file to a base directory, rejects path traversal, and corrects the readOnlyHint annotation to False. Users of the deprecated awslabs.cost-analysis-mcp-server package are told to migrate. If you cannot upgrade, remove generate_cost_report from auto-approve lists.

Jarvis Osorio (Jarvis Nelson Osorio) portrait in red and pink neon light
Jarvis Osorio (Jarvis Nelson Osorio) in red and pink light.

What I am changing in my own practice I wrote mcp-pordego to scan MCP setups in Claude Desktop, Claude Code, Cursor and VS Code for unpinned npx, uvx, Docker and git sources, inline secrets, plain-HTTP remotes, broad filesystem roots and shell-capable servers. This week's advisories sharpen that kind of scan: a readOnlyHint can be wrong, a Git MCP server can be dangerous on an untrusted .git/config, a workbench can be crashed by the server it inspects, and a gateway can ignore the switch you thought you set.

agentscrub, the Rust CLI in the agent-session-scrubber repository, exists because agent transcripts accumulate secrets and PII. After a week like this, I would rather have those transcripts redacted by default, dry-run first, than discover later that a crashed workbench or a mis-routed identity header also captured a token. Project page: https://jarvisosorio.com/agent-session-scrubber/ Repository: https://github.com/JarvisOsorio/agent-session-scrubber

Enterprise buyers will hear a related story later this week at Microsoft Ignite. Microsoft's 10 November 2026 Agent 365 guide lists a 19 November lightning talk, LTG1238, on managing MCP servers, plugins and skills — discovering, registering and governing agent tools, and blocking unsanctioned ones. Guide: https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750

If you run these components, the sourced actions are: mcp-server-git 2026.11.11 or later; Inspector 2.15.0; ContextForge 1.1.1, or identity propagation off; AWS Pricing MCP Server 1.0.31, and do not auto-approve generate_cost_report until you have upgraded. A weekly standards-and-releases roundup is at /blog/jarvis-nelson-osorio-agentic-ai-week-november-17-2026/.

Sources - https://github.com/modelcontextprotocol/servers/security/advisories/GHSA-6hqx-gqrx-c5r2 - https://github.com/modelcontextprotocol/inspector/security/advisories/GHSA-9w48-x22j-rrm7 - https://www.ibm.com/support/pages/security-bulletin-ibm-contextforge-mcp-gateway-affected-information-exposure-vulnerability - https://github.com/awslabs/mcp/security/advisories/GHSA-mqm3-w68j-75mw - https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750 - https://jarvisosorio.com/mcp-pordego/ - https://jarvisosorio.com/agent-session-scrubber/

Links