Skip to content

CVE-2026-116066: Firecrawl (< 2.11.586) Playwright Service SSRF Protection Bypass via DNS Rebinding & Companion Agentic AI / GEO Flaws in Leon & NiubiGEO (CVE-2026-116073, CVE-2026-116069)

Density:
C-Level Executive Briefing CVE-2026-116066 CRITICAL SCORE 95/100

Critical agentic AI web-scraping, personal assistant model downloader, and Generative Engine Optimization (GEO) vulnerability cluster flagged by VulnCheck Initial Access Intelligence on October 26, 2026. Under CVE-2026-116066 (CVSS 8.6 / VulnCheck Weaponized 9.5), Mendable.ai Firecrawl prior to version 2.11.586 validates target URLs against private and link-local IP ranges using a pre-flight DNS lookup (isInternalHost) before passing the original hostname to the Playwright headless Chromium rendering service. Because Chromium performs an independent second DNS resolution at socket connection time, an attacker operating an authoritative DNS server with TTL=0 can execute a Time-of-Check to Time-of-Use (TOCTOU) DNS rebinding attackโ€”returning a benign public IP during pre-flight validation and 169.254.169.254 (cloud metadata) or 127.0.0.1 during Playwright navigation, exfiltrating cloud IAM credentials directly into the LLM/RAG Markdown response. Dissected alongside Leon AI (< commit 50df142) percent-encoded path traversal arbitrary file write in getModelFileNameFromURL() via the /download model command (CVE-2026-116073) and NiubiGEO (< 0.3.2) unauthenticated SSRF via the /audit-plan domain field (CVE-2026-116069).

HYBRID

HERMES THREAT SCORE & AGENTIC AI / RAG INGESTION EVALUATION

Target: Firecrawl Playwright Service SSRF Proxy (isInternalHost), Leon AI getModelFileNameFromURL() (/download model) & NiubiGEO /audit-plan Endpoint
Confidence: 99%
95 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 19 / 20
Weaponization 20 / 20
Exposure 19 / 20
Prevalence 19 / 20
Impact 18 / 20
Exploit Maturity 19 / 20
Attack Chain Potential 20 / 20
โš–๏ธ Divergence & Operational Rationale

Standard CVSS v3.1 metrics score CVE-2026-116066 at 8.6 (High) focusing on read-only SSRF impact, whereas VulnCheck evaluates this weaponized agentic web-scraping and AI model-loader cluster at 9.5 (Critical). The Hermes Threat Score assigns 95 (CRITICAL) with a Hermes Agentic Security Score (HASS) of 97 (EXTREME). Firecrawl is the de facto headless-browser ingestion backbone for autonomous RAG pipelines, Model Context Protocol (MCP) web-search tools, and deep-research agents. Because Firecrawl (< 2.11.586) validates the hostname via isInternalHost() in a pre-flight DNS query before handing the raw URL to Chromium/Playwrightโ€”which performs an independent second DNS lookupโ€”an attacker-controlled authoritative DNS server (TTL=0) trivially rebinds the target to 169.254.169.254 or internal Kubernetes microservices, returning cloud IAM credentials directly in the Markdown scrape response. Paired with Leon AI's percent-encoded path traversal in getModelFileNameFromURL() during /download model execution (CVE-2026-116073) and NiubiGEO's unauthenticated /audit-plan SSRF (CVE-2026-116069), this cluster enables zero-click cloud credential theft and arbitrary file overwrite across agentic AI infrastructure.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-116066: Firecrawl (< 2.11.586) Playwright Service SSRF Protection Bypass via DNS Rebinding & Companion Agentic AI / GEO Flaws in Leon & NiubiGEO (CVE-2026-116073, CVE-2026-116069)VULNERABILITY

Connected Nodes: 3
Active Relationships (Outgoing)
→ affectsPRODUCTMendable.ai Firecrawl AI Web Scraping Engine
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in Mendable.ai Firecrawl AI Web Scraping Engine documented in Hermes dossier.”

Supporting Verified Evidence:
→ exploitsAGENTIC ATTACK_PATTERNAAP-001: Direct System Prompt Override
92% VERY_HIGH

Adversary injects explicit formatting tags or role-inversion prompts directly into user input to strip system instructions and escape developer-enforced guardrails.

๐Ÿ” Why is this related? (Evidence & Provenance)

“CVE-2026-116066 weaponizes the agentic attack pattern formalized under AAP-001.”

Supporting Verified Evidence:
→ exploitsAGENTIC ATTACK_PATTERNAAP-003: Tool Parameter Tampering & Built-in Bypass
92% VERY_HIGH

Adversarial subversion of structured tool execution arguments (SQL, Shell, Filepath) passed from an LLM agent to host OS tools or MCP endpoints.

๐Ÿ” Why is this related? (Evidence & Provenance)

“CVE-2026-116066 weaponizes the agentic attack pattern formalized under AAP-003.”

Supporting Verified Evidence:

On October 26, 2026, VulnCheck Initial Access Intelligence and the National Vulnerability Database (NVD) published three high-priority weaponized advisories targeting AI-native headless web scraping engines (Mendable.ai Firecrawl), open-source AI personal assistants (Leon AI), and Generative Engine Optimization (GEO) AI visibility platforms (NiubiGEO):

  1. CVE-2026-116066 (CWE-918 / CWE-350 / CWE-367, CVSS 8.6 High / VulnCheck 9.5 Weaponized โ€” Firecrawl < 2.11.586 SSRF Protection Bypass via DNS Rebinding in Playwright Service): Firecrawl (mendableai/firecrawl), the foundational open-source and self-hosted web-to-Markdown scraping engine powering LangChain, LlamaIndex, Dify, and Model Context Protocol (MCP) browser tools, attempts to block Server-Side Request Forgery (SSRF) by resolving user-supplied URLs via isInternalHost() before launching a headless browser session. However, in versions prior to 2.11.586, the Playwright service (apps/playwright-service-ts) performs this DNS validation at the application layer and then passes the original unpinned URL string to Chromium (page.goto(url)). Using a DNS rebinding attack (TTL=0), an attacker returns a safe public IP during isInternalHost() and immediately switches the DNS A record to 169.254.169.254 (AWS/GCP/Azure IMDS), 127.0.0.1, or internal VPC endpoints when Playwright connectsโ€”returning internal secrets rendered as clean Markdown to the calling AI agent (AAP-001 / AAP-003).
  2. CVE-2026-116073 (CWE-22, CVSS 8.8 / VulnCheck 9.5 Weaponized โ€” Leon AI < commit 50df142 Arbitrary File Write via /download model Path Traversal): In Leon (leon-ai/leon), an open-source personal AI assistant, the built-in /download model <url> chat/CLI command extracts the destination filename from the remote URL using getModelFileNameFromURL(). Prior to commit 50df142 (50df142a89a0859e5b7283d87afdab7b01e712a2), getModelFileNameFromURL() URL-decoded the last path segment without stripping ../ traversal sequences or applying path.basename(). A crafted model URL containing percent-encoded traversal (%2e%2e%2f) writes the downloaded payload outside the intended models directory to any filesystem path writable by the Leon process.
  3. CVE-2026-116069 (CWE-918, CVSS 8.6 / VulnCheck 9.5 Weaponized โ€” NiubiGEO < 0.3.2 Unauthenticated SSRF via /audit-plan domain Field): In NiubiGEO (an AI brand visibility and Generative Engine Optimization auditing tool), the unauthenticated /audit-plan API endpoint accepts a domain field and initiates server-side HTTP fetches without restricting IP literals, localhost, RFC 1918 private subnets, or link-local metadata addresses (169.254.169.254).

1. CVE-2026-116066: Firecrawl (< 2.11.586) TOCTOU DNS Rebinding in Playwright Service

Section titled โ€œ1. CVE-2026-116066: Firecrawl (< 2.11.586) TOCTOU DNS Rebinding in Playwright Serviceโ€

When an AI agent or API client submits a scrape request (POST /v1/scrape) to a self-hosted Firecrawl deployment running version < 2.11.586, the request is routed to the Playwright rendering microservice. To prevent attackers from scraping internal infrastructure, the service implements an SSRF guard (isInternalHost) that resolves the hostname and checks whether the resulting IPv4/IPv6 address falls within private, loopback, or link-local ranges:

// VULNERABLE PATTERN: Mendable.ai Firecrawl < 2.11.586 (CVE-2026-116066)
// Component: Playwright Service SSRF Guard & Navigation Handler
import dns from 'node:dns/promises';
import ipaddr from 'ipaddr.js';
async function isInternalHost(hostname: string): Promise<boolean> {
// STEP 1 (Time-of-Check): Node.js resolves the hostname via OS/c-ares DNS lookup.
// Attacker's authoritative DNS server responds with 93.184.216.34 (Public IP, TTL=0).
const { address } = await dns.lookup(hostname);
const parsed = ipaddr.parse(address);
const range = parsed.range();
return ['private', 'loopback', 'linkLocal', 'uniqueLocal', 'unspecified'].includes(range);
}
export async function scrapeWithPlaywright(url: string) {
const parsedUrl = new URL(url);
// Pre-flight check passes because the first DNS query returned 93.184.216.34!
if (await isInternalHost(parsedUrl.hostname)) {
throw new Error('SSRF Blocked: Internal host access is forbidden');
}
// STEP 2 (Time-of-Use โ€” CWE-367 / CWE-918):
// Chromium / Playwright navigates to the raw `url` string (`http://rebind.attacker.com/latest/meta-data/`).
// Because TTL=0, Chromium's network service performs a SECOND DNS lookup, which now resolves
// to 169.254.169.254 (or 127.0.0.1:6379 / internal Kubernetes service) without re-running isInternalHost()!
const page = await browserContext.newPage();
await page.goto(url, { waitUntil: 'networkidle' });
return await page.content();
}

In version 2.11.586, Firecrawl hardened its Playwright service proxy layer to validate resolved IP addresses at the proxy/interception boundary rather than relying solely on a pre-navigation check, and documented that self-hosted deployments should enforce an egress PROXY_SERVER blocking RFC 1918 and link-local destinations at the network socket level.

2. CVE-2026-116073: Leon AI (< commit 50df142) Percent-Encoded Path Traversal in /download model

Section titled โ€œ2. CVE-2026-116073: Leon AI (< commit 50df142) Percent-Encoded Path Traversal in /download modelโ€

In Leon AI (leon-ai/leon), users and skills can download LLM weights or voice models via the /download model <url> command. The helper function getModelFileNameFromURL() in the model downloader extracted the filename from the URL path and decoded percent-encoded characters after splitting on /:

// VULNERABLE PATTERN: Leon AI < commit 50df142 (CVE-2026-116073)
// Component: Model Download Utility (getModelFileNameFromURL)
import path from 'node:path';
import fs from 'node:fs';
function getModelFileNameFromURL(modelUrl: string): string {
const parsedUrl = new URL(modelUrl);
// Splits on literal '/' first: the last segment is "%2e%2e%2f%2e%2e%2f.ssh%2fauthorized_keys"
const segments = parsedUrl.pathname.split('/');
const rawFileName = segments[segments.length - 1];
// FLAW (CWE-22): decodeURIComponent() converts "%2e%2e%2f" into "../" AFTER the '/' split!
// No path.basename() or canonical path containment check is performed afterwards.
return decodeURIComponent(rawFileName);
}
async function downloadModel(modelUrl: string, modelsDir: string) {
const fileName = getModelFileNameFromURL(modelUrl);
// Resolves outside `modelsDir` to arbitrary writable filesystem paths!
const destinationPath = path.join(modelsDir, fileName);
const writer = fs.createWriteStream(destinationPath);
// ... streams attacker-controlled HTTP response body to destinationPath ...
}

In commit 50df142 (50df142a89a0859e5b7283d87afdab7b01e712a2), Leon AI sanitized the extracted filename with path.basename() after URI decoding and verified that the resolved path stays strictly inside the designated models directory.

3. CVE-2026-116069: NiubiGEO (< 0.3.2) Unauthenticated /audit-plan SSRF

Section titled โ€œ3. CVE-2026-116069: NiubiGEO (< 0.3.2) Unauthenticated /audit-plan SSRFโ€

In NiubiGEO prior to 0.3.2, the /audit-plan endpoint accepts an unauthenticated JSON request containing a domain field (intended to fetch a target websiteโ€™s robots.txt, llms.txt, and homepage for AI visibility scoring). Because the backend concatenates or parses domain directly into an outbound HTTP client without authentication or IP range filtering, supplying "domain": "http://169.254.169.254/latest/meta-data/iam/security-credentials/" or "domain": "127.0.0.1:8080/admin" allows unauthenticated attackers to probe loopback services and cloud metadata endpoints.

flowchart TD
subgraph FirecrawlRebind["CVE-2026-116066: Firecrawl (< 2.11.586) Playwright DNS Rebinding SSRF"]
F1["Attacker Submits URL (rebind.attacker.com) to Firecrawl Scrape API / MCP Tool"]
F2["Lookup 1 (isInternalHost): Authoritative DNS Returns 93.184.216.34 (TTL=0) -> PASS"]
F3["Lookup 2 (Playwright Chromium page.goto): DNS Rebinds to 169.254.169.254 / 127.0.0.1"]
F4["Cloud IAM Metadata / Internal VPC Response Returned as Markdown to Agent"]
F1 --> F2 --> F3 --> F4
end
subgraph LeonNiubi["CVE-2026-116073 (Leon AI) & CVE-2026-116069 (NiubiGEO < 0.3.2)"]
L1["CVE-2026-116073: /download model URL with %2e%2e%2f<br/>getModelFileNameFromURL() Decodes '../' -> Arbitrary File Write"]
N1["CVE-2026-116069: Unauthenticated POST /audit-plan with Internal domain<br/>Direct Server-Side Fetch to Loopback & Cloud IMDS"]
end
style F3 fill:#ff1a1a,stroke:#800000,color:#ffffff
style F4 fill:#ff4d4d,stroke:#990000,color:#ffffff
style L1 fill:#ff1a1a,stroke:#800000,color:#ffffff
style N1 fill:#ff6600,stroke:#993300,color:#ffffff

To detect exploitation of Firecrawl Playwright DNS rebinding (CVE-2026-116066), Leon AI /download model path traversal (CVE-2026-116073), and NiubiGEO /audit-plan SSRF (CVE-2026-116069), deploy the following multi-layered detection suite:

title: Agentic AI SSRF & Path Traversal in Firecrawl Playwright, Leon AI & NiubiGEO (CVE-2026-116066 / CVE-2026-116073 / CVE-2026-116069)
id: 116066a1-2026-4c88-9b11-firecrawlleon26
status: experimental
description: Detects headless Chromium/Playwright processes initiated by Firecrawl connecting to link-local cloud metadata (169.254.169.254), Leon AI /download model percent-encoded path traversal, or NiubiGEO /audit-plan SSRF probes.
references:
- https://vulncheck.com/advisories/firecrawl-before-2.11.586-ssrf-protection-bypass-via-dns-rebinding-in-playwright-service
- https://vulncheck.com/advisories/leon-before-commit-50df142-path-traversal-via-download-model-command
- https://vulncheck.com/advisories/niubigeo-before-0.3.2-unauthenticated-ssrf-via-audit-plan-domain-field
author: Hermes Threat Intelligence
date: 2026/10/26
logsource:
category: network_connection
product: linux
detection:
selection_playwright_imds:
Image|endswith:
- '/chrome'
- '/chromium'
- '/headless_shell'
DestinationIp:
- '169.254.169.254'
- '100.100.100.200'
- '127.0.0.1'
condition: selection_playwright_imds
falsepositives:
- Local healthcheck probes specifically targeting a known loopback port
level: critical
tags:
- attack.initial_access
- attack.credential_access
- attack.t1552.005
- attack.t1190
- cve.2026-116066
- cve.2026-116073
- cve.2026-116069

  1. Upgrade Firecrawl to >= 2.11.586 & Enforce Egress Proxy Pinning (CVE-2026-116066):
    • Upgrade all self-hosted Firecrawl instances and Playwright microservice containers to version 2.11.586 or later.
    • Configure the PROXY_SERVER environment variable for the Playwright service to route all headless Chromium traffic through a hardened egress proxy (such as egress-filter or Envoy/Squid with DNS pinning) that blocks 169.254.169.254/32, 127.0.0.0/8, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, and fc00::/7 at the TCP connection socket level, and enforce IMDSv2 (HttpPutResponseHopLimit = 1) on all cloud worker nodes.
  2. Patch Leon AI Model Downloader (CVE-2026-116073):
    • Update Leon AI (leon-ai/leon) to commit 50df142 (50df142a89a0859e5b7283d87afdab7b01e712a2) or later so that getModelFileNameFromURL() applies path.basename() after decodeURIComponent() and validates destination path containment before writing model weights to disk.
  3. Upgrade NiubiGEO to >= 0.3.2 (CVE-2026-116069):
    • Upgrade NiubiGEO to version 0.3.2 or later, require authentication on /audit-plan, and enforce strict IP-level SSRF filtering on the resolved socket address.