registry / Credential exfiltration / @finchagentic/mcp
@finchagentic/mcp
The runtime layer for Agentic AI. Persistent memory, autonomous agents, and workflows that survive every session.
This report shows the evidence for @finchagentic/mcp@4.7.5. 2 of 10 known versions are blocked or flagged. See all versions.
AI Security Review
scanned 2d ago · by lpm-firewall-aiThe shared Finch backend client automatically forwards configured third-party provider keys on ordinary tool requests, not only on requests that need those keys for the documented reasoning features. This is credential exfiltration to the package vendor without clear, scoped consent.
Static reason
No blocking static signals were detected.; previous stored version diff introduced dangerous source
Trigger
A user invokes a tool that calls the Finch backend while one or more provider API keys are configured in the environment.
Impact
Finch's backend receives credentials for third-party AI services even on requests unrelated to the documented server-side reasoning features.
Mechanism
The shared backend client unconditionally copies configured Anthropic, OpenAI, Grok, and Bankr keys into request headers sent to the Finch Convex endpoint.
Attack narrative
When a backend-backed tool runs, the shared client sends any configured Anthropic, OpenAI, Grok, or Bankr API keys to Finch's Convex endpoint. This happens for ordinary operations such as listing memory, although the README describes those keys as needed for specific server-side reasoning features. The keys are therefore disclosed to the package vendor beyond the stated feature context.
AI rationale
The shared Finch backend client automatically forwards configured third-party provider keys on ordinary tool requests, not only on requests that need those keys for the documented reasoning features. This is credential exfiltration to the package vendor without clear, scoped consent.
Evidence
dist/convex.jsdist/tools/memory.jsREADME.md
Network endpoints1
valuable-fish-533.convex.site
Decision evidence
public snapshotAI called this Malicious at 97.0% confidence as Malware with low false-positive risk.
Evidence for block
- The backend client defaults to https://valuable-fish-533.convex.site.
- Every Finch backend request attaches any configured Anthropic, OpenAI, Grok, and Bankr API keys as headers.
- Ordinary memory operations call the same backend client, so key forwarding is not limited to LLM reasoning requests.
- The README describes these keys as needed for specific server-side reasoning features, not as credentials to include with all backend requests.
Evidence against
- The README describes the keys as BYOK for server-side reasoning, indicating a stated product use for them.
- The forwarding occurs on backend requests triggered by tool calls, rather than during package installation.
Behavioral surface
ChildProcessCryptoEnvironmentVarsFilesystemNetwork
HighEntropyStringsUrlStrings
Source & flagged code
5 flaggeddist/local-vault.jsView file
61// "not found", matching callConvex's own 404 format.
L62: const VAULT_DIR = path.join(os.homedir(), ".finch", "vault");
L63: // Local vault is enabled purely by config - it needs no server, just the
...
L82: const raw = fs.readFileSync(indexPath(cfg), "utf8");
L83: const parsed = JSON.parse(raw);
L84: return { entries: parsed.entries ?? {}, links: parsed.links ?? [] };
...
L168: agentId: input.agentId, tags, pinned: false, currentVersion: 1,
L169: createdAt: ts, updatedAt: ts, metadata: input.metadata,
L170: versions: [{ version: 1, commitMsg: input.commitMsg, agentId: input.agentId, size: Buffer.byteLength(input.content), createdAt: ts }],
Low
Weak Crypto
Package source references weak cryptographic algorithms.
dist/local-vault.jsView on unpkg · L61dist/agent-loop.jsView file
30async function runAgent(userMessage, history, onToolCall) {
L31: const provider = process.env.FINCH_PROVIDER?.toLowerCase().trim();
L32: const bankrKey = process.env.BANKR_API_KEY;
...
L144: const sendTurn = async (messages, tools) => {
L145: const res = await fetch("https://api.anthropic.com/v1/messages", {
L146: method: "POST",
...
L151: },
L152: body: JSON.stringify({ model, max_tokens: 4096, system: SYSTEM_PROMPT, tools, messages }),
L153: signal: AbortSignal.timeout(90000),
...
L155: if (!res.ok) {
L156: const body = await res.text().catch(() => "");
L157: throw new Error(`Anthropic ${res.status}: ${body.slice(0, 300)}`);
Findings
1 Critical3 High2 Medium6 Low
CriticalPrevious Version Dangerous Deltadist/cli-setup.js
HighCredential Exfiltrationdist/agent-loop.js
HighTrigger Reachable Credential Exfiltrationdist/agent-loop.js
HighSemantic Analysis Limiteddist/agent-loop.js#virtual:normalized:round1
MediumNetwork
MediumEnvironment Vars
LowNon Install Lifecycle Scripts
LowScripts Present
LowWeak Cryptodist/local-vault.js
LowFilesystem
LowHigh Entropy Strings
LowUrl Strings
Affected versions and remediation
This report applies to @finchagentic/mcp@4.7.5.
The versions list below shows the verdict for each recorded version.
- Avoid installing @finchagentic/mcp@4.7.5. Remove it from direct dependencies and check your lockfile for transitive copies.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
- If this version ran, investigate the affected machine and build environment. Rotate credentials it could access and rebuild from a trusted environment.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.