CVE-2026-54353: Budibase: SSRF via DNS rebinding in automation steps

Summary Authenticated users with automation permissions can bypass Budibase's SSRF blacklist through DNS rebinding. The outbound fetch flow validates a hostname against the blacklist before the request is sent, but the actual socket connection later performs a separate DNS lookup through node-fetch. Since the validated IPs are never pinned to the connection, an attacker-controlled hostname can return a public IP during validation and a private/internal IP during the real connection. This results in a non-blind SSRF primitive against internal services reachable from the Budibase host, including loopback, RFC1918 ranges, and cloud metadata endpoints. Details The issue comes from the outbound fetch validation flow resolving DNS twice: During blacklist validation Again during the real socket connection The first lookup result is discarded after validation, so the second lookup is free to resolve to a different IP. This creates a classic TOCTOU DNS rebinding issue. Affected flow in: packages/backend-core/src/utils/outboundFetch.ts ``` async function throwIfUnsafe(url: string): Promise<void> { const parsed = parseUrl(url) if (await isBlacklisted(parsed.hostname)) { throw new Error("URL is blocked or could not be resolved safely.") } } for (let redirects = 0; redirects <= MAX_REDIRECTS; redirects++) { await throwIfUnsafe(nextUrl) const response = await fetchFn(nextUrl, nextRequest) // ... } ``` fetchFn uses plain node-fetch with no custom http.Agent / https.Agent, so the underlying socket performs its own independent dns.lookup after validation completes. The same pattern also exists in: packages/server/src/automations/steps/utils.ts ``` await throwIfBlacklisted(nextUrl) const response = await fetch(nextUrl, nextRequest) ``` The blacklist implementation resolves hostnames but only returns a boolean: packages/backend-core/src/blacklist/blacklist.ts ``` async function lookup(address: string): Promise<string[]> { address = parseAddress(address) const addresses = await performLookup(address, { all: true }) return addresses.map(addr => addr.address) } export async function isBlacklisted(address: string): Promise<boolean> { // ... if (!net.isIP(address)) { try { ips = await lookup(address) } catch (e) { /* ... */ } } else { ips = [address] } return ips.some(ip => blackList!.check(ip, getIpVersion(ip))) } ``` The resolved IPs are discarded, so callers cannot pin the later socket connection to the validated addresses. An attacker controlling authoritative DNS for a hostname can therefore return: a public IP during validation a private/internal IP during the actual connection Anything routing through these helpers inherits the issue, including: outgoing webhook Slack Discord Make Zapier n8n AI extract object-store fetches Several of these steps return upstream response content directly into automation output, which makes the SSRF non-blind. PoC Tested locally against a self-hosted build from master. No Budibase-operated infrastructure was touched. Run Budibase locally. Start a harmless local HTTP listener: python3 -m http.server 8080 --bind 127.0.0.1 Use a rebinding hostname such as: 7f000001.cb007264.rbndr.us which rotates between: 127.0.0.1 203.0.113.100 Steps to reproduce: Log into Budibase with automation permissions. Create an automation using the Outgoing Webhook step. Set the URL to: http://<rebinding-host>:8080/ Trigger the automation. Observed result: The blacklist validation resolves the hostname to the public IP and allows the request. node-fetch performs a second DNS lookup during socket creation. The second lookup resolves to 127.0.0.1. The TCP connection lands on the local service. The local server response body appears directly in the automation output. Impact This produces a non-blind read-SSRF primitive against anything reachable from the Budibase host process, including: loopback services (127.0.0.1) RFC1918 ranges internal Kubernetes/VPC services cloud metadata endpoints (169.254.169.254) On cloud deployments without IMDSv2 enforcement, this may expose temporary IAM credentials via: /latest/meta-data/iam/security-credentials/<role> On multi-tenant hosted deployments, this may also create potential cross-tenant access paths through shared internal infrastructure.

An attacker with a standard Budibase account holding automation permissions registers a domain and configures its authoritative DNS to alternate between a public IP (e.g., 203.0.113.100) and the target internal IP (e.g., 169.254.169.254 for cloud metadata). They create a Budibase automation with an Outgoing Webhook step targeting their rebinding domain on port 80. When triggered, throwIfUnsafe() resolves the hostname — DNS returns the public IP, the blacklist check passes, and the resolved addresses are discarded. Milliseconds later, node-fetch performs its own independent DNS lookup during socket creation; the attacker's DNS now returns 169.254.169.254. The TCP connection lands on the AWS metadata service, and the response body — including IAM role credentials at /latest/meta-data/iam/security-credentials/<role> — appears directly in the automation output and is visible to the attacker, enabling lateral movement via stolen temporary credentials.