Malicious npm packages posing as AI SDKs and MCP servers: character-ai-sdk, char-ai-sdk, @mibbp/agent-observer, mrnima-openai-chatgpt-bot and @fdcasj_dfew/md-doc-mcp
Between 18 and 19 November 2026, OSV and GitHub flagged five more AI-themed npm packages as malicious. Two pose as AI persona/memory SDKs (character-ai-sdk, char-ai-sdk) and load large obfuscated blobs as soon as they are imported. @fdcasj_dfew/md-doc-mcp is a Markdown MCP server that probes the Alibaba Cloud metadata service at startup, a republish of the @ifasn doc-MCP packages. @mibbp/agent-observer and mrnima-openai-chatgpt-bot are GitHub malware advisories. Treat any machine that installed or ran one as compromised.
What it is
character-ai-sdk 1.0.0, 1.0.1 (MAL-2026-20196) and char-ai-sdk 1.0.0 (MAL-2026-20195): advertised as AI persona/memory SDKs (char-ai-sdk mentions Bedrock and OpenAI). The main entry re-exports personas.mjs, which imports ./cache-warm.mjs, an obfuscator.io-style blob of about 300 KB, with sibling blobs build-cache.mjs, cache-sweep.mjs and workspaces.js. Decoded strings reference process control, hostname and child-process APIs. character-ai-sdk also ships native cache-probe.node addons built from a VM-detection library (anti-sandbox). Reported by Amazon Inspector.
@fdcasj_dfew/md-doc-mcp 1.0.0, 1.0.1 (MAL-2026-20187): a Markdown toolkit MCP server (md_toc, md_stats, md_table) whose index.js calls fieldCheck() at load, before any MCP message. It requests the Alibaba Cloud instance metadata service (100.100.100.200) for hostname, instance-id, vpc-id, private IPv4 and region, and writes the results to stderr and output/field_check.txt. It matches the earlier @ifasn/md-doc-mcp, @ifasn/mmdd-doc-mcp and @ifasn/ms-doc-mcp packages republished under a new account.
@mibbp/agent-observer (GHSA-2r45-7hg7-w99r / MAL-2026-20254, 19 November) and mrnima-openai-chatgpt-bot (GHSA-xvx7-m3mm-3jpw / MAL-2026-20155, 19 November): GitHub malware advisories, all versions. They don't describe the payload; GitHub's guidance is that any computer that installed or ran them should be considered fully compromised.
MCP servers and agent add-ons are often started with npx straight from a client config, so one launch is enough to run the code with your user's access to files, SSH keys, cloud credentials and API keys.
How to check
- Search lockfiles, global installs, the npx cache and MCP client configs (Claude Desktop, Cursor, VS Code, Windsurf) for these five names.
- On cloud VMs where md-doc-mcp ran, look for an output/field_check.txt file and review instance metadata access and RAM role credential use.
How to fix
- Remove the packages and their MCP config entries, and clear the npx cache.
- From a clean machine, rotate every secret that was reachable on an affected one: LLM API keys, cloud credentials (including instance RAM role credentials), GitHub/npm tokens, SSH keys. Rebuilding the machine is the safe option.
- Install AI SDKs and MCP servers only from the vendor's documented package name, and pin versions.
grep -rnE '"(character-ai-sdk|char-ai-sdk|@mibbp/agent-observer|mrnima-openai-chatgpt-bot|@fdcasj_dfew/md-doc-mcp)"' package-lock.json pnpm-lock.yaml yarn.lock 2>/dev/null
grep -rlE 'character-ai-sdk|char-ai-sdk|agent-observer|mrnima-openai|md-doc-mcp' ~/.npm/_npx 2>/dev/null
find / -name field_check.txt -path '*output*' 2>/dev/nullSources
- criticalMalicious npm MCP servers @ifasn/md-doc-mcp, mdd-doc-mcp, mmd-doc-mcp, mmdd-doc-mcp and ms-doc-mcp probe cloud metadata for credentials
- criticalMalicious npm packages posing as MCP servers and AI APIs: @shadowdu/memos-mcp, @shadowdqj/image-mcp, mem0-mcp-server and ai-apis
- criticalMalicious AI agent and MCP packages: mcp-aevo-server (PyPI) steals Aevo trading keys and wallet private keys; vamp-mcp, tsc-ai and @tnhorg/agent-skills (npm) flagged as malware
- criticalMalicious npm package @guicai777/mcp-doctor posed as an MCP diagnostic tool and dropped a hidden Windows executable, with a README telling users to run mshta (MAL-2026-20295)
- criticalCVE-2026-77359: Jupyter MCP Server's /api/connect and /api/stop routes skipped the token and Origin checks, so a website or network caller could repoint it to an attacker's Jupyter backend (fixed 1.0.3)
- criticalCVE-2026-76807: Doist's official Todoist MCP HTTP server let any website use DNS rebinding to read, create and delete your Todoist data (fixed 10.3.3)
- criticalMalicious npm and PyPI packages posing as AI and MCP tools: depfix-ai, agent-trace-kit, @go-gost/mcp-darwin-x64 and my-mcp-server-driving-license-ocr