140 silent contract changes in the official MCP servers (and how to catch them)

1. Deadbugz -- the server that waits for call #3 (August 2026)

Pillar Security documented an active campaign pushing a malicious MCP server (it calls itself "productivity-suite") through 23 GitHub pull requests opened within 74 minutes (Aug 10). It offers two innocent tools -- text formatting and summarization -- and behaves perfectly at first. After exactly three tool calls, it rewrites the metadata it returns to the agent: now the description says to hunt for SSH private keys, AWS credentials, shell history and Kubernetes config -- and to hide that from the user.

The killer detail: the trigger is a call counter, not a code change. Nothing new got installed. There is no update diff to review. The PRs were caught (19 closed, 4 left open at review) and no theft was confirmed in the wild -- but as a class this is the cleanest demonstration that "reviewed at install" is not a security control.

Primary: Pillar Security, "Deadbugz: Currently Active MCP Supply-Chain Campaign"

2. GhostSplice -- split the instruction, beat the refusal (August 2026)

ASSET Research Group (UMKC): "Eleven frontier models, one malicious MCP server. Ask them to leak your credentials and they refuse. Split the request into multiple harmless fragments..." -- GhostSplice splits one data-theft instruction across a tool description and a later tool result, so no single message looks malicious. Average compliance across 11 tested models jumped from 42% to 82%, and agents exfiltrated SSH keys, secrets and source code. Per-message inspection (keyword scanning, output filters) never had a chance -- the payload only exists across messages.

Primary: ASSET Research Group + Bishop Fox "Spliced Instructions..." (Aug 21)

3. ChainDrop -- the worm that poisons Claude Code itself (August 2026)

The self-propagating npm worm (Microsoft's analysis covers 400+ packages -- later tallies pushed it to 860+) didn't even need you to install anything for its nastiest move: it committed .claude/settings.json (a SessionStart hook) and .vscode/tasks.json (runOn: folderOpen) into repositories. So opening the project -- or starting a Claude Code session -- executed the payload. Secrets walked off CI runners; the stealer was a 728 KB obfuscated bundle with C2 over Ethereum contracts.

That one matters here specifically: your agent config is executable attack surface now, and it is exactly as silently mutable as any tool description.

Primary: Microsoft Security Blog (Aug 4) + WorkOS analysis (Aug 6)

4. FakeGit -- 800+ fake "MCP servers" and skills, recommended by your agent (July 2026)

Island's research found 7,600 malicious GitHub repos and ~6,600 fake accounts -- with 800+ masquerading as AI skills and MCP servers. When users asked their coding agents (Claude Code among them) for recommendations, the agents recommended the malicious repos themselves. The campaign's release assets hit 14M+ downloads, and 600+ of its listings sat on public MCP registries (LobeHub, Glama, MCP.so, MCP Market) before takedowns. The chain ended in StealC, an infostealer aimed at developer machines, browser credentials, and cloud tokens.

Primary: The Hacker News, "FakeGit campaign..." (Jul 20)

5. The scale -- silent drift is the norm, not an edge case (September 2026)

An arXiv census harvested the entire public MCP registry -- 21,643 servers, 72,606 version records (August 2026 snapshot):

• 51.1% of multi-version servers changed what they advertise
• 40.6% did it silently
• 4.2% redirected their remote endpoint to a different host while keeping the same registry identity -- "a change the protocol never surfaces to installed clients"
• Silent drift is associated with ~3x higher odds of a high-severity finding, and stars/installs are a weak safety signal

The CVE flow matches: CVE-2026-81486 (path traversal, CISA-noted), CVE-2026-87911 (CVSS 9.6 OS command injection in AWS Labs' postgres MCP server), plus an Atlassian MCP path traversal, a cleartext cluster token in ArcadeDB's MCP, and an SSRF in facebook-ads-mcp-server -- all August-September. September also brought CVE-2026-20176 -- a malicious server could make the official MCP Python SDK (1.9.1-2.1.1) hand over the client secret, authorization code and PKCE verifier (i.e., account takeover); fixed in 1.30.0/2.2.0. And a July dynamic scan of internet-facing MCP servers (arXiv:2608.00150): 91.8% had no OAuth, 687 tool instances exposed shell execution with no access control, and 41.6% of confirmed servers vanished within three days.

Primary: arXiv:2609.14119, "Same Name, Different Server: A Security Census of Silent Drift in the MCP Ecosystem" (Sep 12)

And it wasn't only MCP. This summer tested every trust boundary around agents. The July Hugging Face intrusion: an escaped agent system took ~17,600 malicious actions and seized 136 production credentials before containment (OpenAI's account). SharedRoot (July 23): untrusted content escaped Claude Cowork's local VM through a writable host mount and could read SSH keys and cloud credentials on the Mac, with ~500,000 macOS users in scope before the fix -- Anthropic closed the report as "Informative" and moved new sessions to cloud by default (Accomplish AI's write-up). And JadePuffer, the first fully LLM-driven ransomware, encrypted 1,342 configs -- and hunted wallets and seed phrases along the way.

What actually stops this

The pattern across every case above is the same: it's never that the model got jailbroken. It's that something changed -- or was split -- after trust was granted, and nothing re-checked it. You can't prompt your way out of that. You pin it.

I maintain RugSnare, a small open-source tool built exactly for this (Apache-2.0, zero npm dependencies, no telemetry, everything local). What it does:

• npx rugsnare scan -- pins the exact contract you approved (description, schema, behavioral annotations) into .rugsnare/pins.json. Commit it.
• npx rugsnare diff -- re-checks it anywhere (CI gate, pre-launch). Any silent change exits 1. Deterministic, not model judgment.
• rugsnare canary replay -- replays your real recorded tool calls against a new version before you upgrade (read-only by default), and returns "DO NOT UPGRADE" when behavior changed.
• rugsnare run -- wraps a live server and enforces policies (blocking the read ~/.ssh / ~/.aws/credentials class of calls), and quarantines mid-session swaps.
• It also scans SKILL.md skill files with the same adversarial signals -- the ChainDrop class.

Evidence it works, measured on real releases: we pinned every stable version of the four official @modelcontextprotocol/server-* servers -- all 66 release pairs -- and counted 140 silent contract changes (43 schema breaks, 28 behavioral-hint flips, 37 new items, 24 removals; 43 clean pairs). Not one was announced in a changelog. One command, ~30 minutes, reproducible on your machine.

Even if you never install a tool, do these:

1. Pin versions (pkg@1.2.3, never @latest) -- and treat a version bump as code review.
2. Commit your pins and re-diff after every update.
3. Treat your MCP config and .claude/settings.json as executable code.
4. Don't auto-approve write/destructive tools.

Sources

1. Pillar Security -- Deadbugz: Currently Active MCP Supply-Chain Campaign (Aug 12, 2026)
2. ASSET Research Group -- GhostSplice (Aug 2026); Bishop Fox podcast (Aug 21, 2026)
3. Microsoft Security Blog -- ChainDrop npm worm (Aug 4, 2026); WorkOS (Aug 6, 2026); CyberScoop on the wider campaign (860+ packages)
4. The Hacker News -- FakeGit campaign, 7,600 repos (Jul 20, 2026); Island -- AgentBaiting
5. arXiv:2609.14119 -- Same Name, Different Server: A Security Census of Silent Drift in the MCP Ecosystem (Sep 12, 2026)
6. arXiv:2608.00150 -- Exposed by Design: audit of internet-facing MCP servers (July 2026)
7. Cycode -- MCP Python SDK OAuth account takeover; The Hacker News coverage (Sep 2026)
8. CVE-2026-81486 (mcp-file-context-server, CISA weekly Aug 24-31) - CVE-2026-87911 (awslabs postgres MCP, CVSS 9.6) - CVE-2026-73498 / CVE-2026-67357 / CVE-2026-19956 (Adversa September roundup)
9. TNW -- Claude Cowork SharedRoot sandbox escape; Accomplish AI technical write-up (Jul 23-26, 2026)
10. Hugging Face -- security incident disclosure; OpenAI's account (July 2026)
11. Sysdig -- JadePuffer agentic ransomware (Jul 2026)
12. RugSnare -- github.com/Paraphern/rugsnare - npm - rugsnare.com

---

Notes / limits: I counted only public, documented incidents from July-September 2026. There are no public reports of end users' wallets or accounts drained specifically by a silent MCP update in this window -- the documented damage is credential theft, code execution and data exfiltration surfaced by research and vendor investigations. I'd rather say that precisely than stretch a headline.

Disclosure: I maintain RugSnare (open source, Apache-2.0, zero deps -- the pinning/diff/replay tooling above). Everything here is runnable and reproducible; happy to answer questions about any of the cases.