Skip to content

CVE-2026-105139: Obot Agent Platform Authorization Bypass via vMCP Profile Prompts and Resources

Density:
HERMES

HERMES THREAT SCORE & AGENTIC INFRASTRUCTURE EVALUATION

Target: Obot AI Agent Platform โ€” Virtual Model Context Protocol (vMCP) Router & Policy Gateway
Confidence: 98%
90 / 100
CRITICAL

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

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 17 / 20
Weaponization 18 / 20
Exposure 18 / 20
Prevalence 16 / 20
Impact 19 / 20
Exploit Maturity 19 / 20
Attack Chain Potential 20 / 20
โš–๏ธ Divergence & Operational Rationale

Traditional CVSS v3.1 scores this vulnerability at 8.8 (High) as an authenticated horizontal authorization bypass across application components. The Hermes Threat Score assesses it at 90 (CRITICAL) with an Agentic Security Score (HASS) of 95 (CRITICAL). In autonomous AI systems, the Model Context Protocol (MCP) serves as the communication standard connecting LLM agents to sensitive knowledge bases, custom tools, and prompt libraries. Obot introduced virtual MCP (vMCP) profiles to enforce tenant isolation and granular access control. However, access policy filters were enforced solely on tool execution (tools/call), omitting validation on prompts, resources, and resource templates. An authenticated user or subagent matching any vMCP profile can extract proprietary system prompts, internal corporate reasoning templates, and confidential vector store resources belonging to other tenants via shared component connections (MITRE ATLAS AML.T0024 & AML.T0040).

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

CVE-2026-105139: Obot Agent Platform Authorization Bypass via vMCP Profile Prompts and ResourcesVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTObot Autonomous AI Agent Platform
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

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

“Confirmed security vulnerability in Obot Autonomous AI Agent Platform documented in Hermes dossier.”

Supporting Verified Evidence:

Obot enables organizations to build, orchestrate, and host autonomous AI agents. To connect agents to external services, Obot relies on the Model Context Protocol (MCP), exposing servers through a virtualized gateway (vMCP) that manages credentials, permissions, and tenant isolation.

ParameterTechnical SpecificationOperational Details
CVE IdentifierCVE-2026-105139International vulnerability catalog identifier
Affected ProductObot AI Agent PlatformEnterprise autonomous multi-agent orchestration runtime
Vulnerable Componentpkg/gateway/vmcp/proxy.go & pkg/api/handlers/vmcp.govMCP reverse proxy JSON-RPC dispatch controller
Vulnerability ClassCWE-285 / CWE-862 / CWE-639Missing Authorization / Bypassed Access Control
Attack VectorNetwork (/api/vmcp/{profile_id}/* JSON-RPC)Authenticated MCP protocol requests
Remediation ReleaseObot v0.26.2Comprehensive policy validation across all MCP methods
[ Low-Privilege User / Compromised Agent ]
|
1. POST /api/vmcp/{profile_id}
JSON-RPC: {"method": "prompts/list"}
|
v
+-------------------------------------------+
| Obot vMCP Gateway Router |
| - Checks Tool Call Authorization? |
| -> NO (Request is 'prompts/list') |
| - Authorization Filter Skipped! |
+-------------------------------------------+
|
v
+-------------------------------------------+
| Shared Component Connection Pool |
| - Uses Profile Owner's Full Privileges |
| - Queries Upstream MCP Servers |
+-------------------------------------------+
|
v
+-------------------------------------------+
| Confidential Enterprise Assets |
| - Tenant System Prompts & Guardrails |
| - Proprietary Financial Contexts |
| - Restricted Resource Templates |
+-------------------------------------------+
|
v
2. Attacker exfiltrates cross-tenant data!

The vulnerability lies in the JSON-RPC request dispatcher implemented in pkg/gateway/vmcp/proxy.go. When processing incoming requests forwarded from an agent LLM, the handler inspects the JSON-RPC method to decide whether access control rules should be evaluated:

// Vulnerable Implementation in pkg/gateway/vmcp/proxy.go
func (p *VMCPProxy) HandleRPC(ctx context.Context, profile *types.VMCPProfile, req *jsonrpc.Request) (*jsonrpc.Response, error) {
// Only tools/call was subjected to profile policy enforcement
if strings.HasPrefix(req.Method, "tools/") {
if err := p.enforceToolPolicy(ctx, profile, req); err != nil {
return nil, err
}
}
// Missing check: prompts/* and resources/* were forwarded directly!
// The request executes through the component owner's shared client connection
conn, err := p.getSharedComponentConnection(profile.ComponentID)
if err != nil {
return nil, err
}
return conn.Forward(ctx, req)
}

Because strings.HasPrefix(req.Method, "tools/") was the sole gatekeeper, requests using the Model Context Protocol standard methods:

  • prompts/list
  • prompts/get
  • resources/list
  • resources/read
  • resources/templates/list

bypassed the authorization layer completely. The gateway used the shared connection established by the profile owner (often an administrative service identity) to query the upstream MCP server, returning unrestricted results to the unauthorized caller.


In modern agentic systems, MCP resources and prompts represent the intellectual property and boundary controls of the agent:

  1. System Prompt Extraction (Prompt Theft): An attacker issues a prompts/get request targeting an internal agent (e.g., executive-finance-advisor). The response discloses the complete system prompt, including proprietary business logic, few-shot demonstration data, and confidential executive guidance.
  2. Knowledge Base Resource Extraction: An attacker queries resources/read with a URI pointing to internal documentation or vector index fragments (internal://knowledge/q3-earnings-draft.md). The vMCP proxy retrieves and returns the text payload without validating workspace access.
  3. Indirect Prompt Injection & Guardrail Evasion: By analyzing extracted prompt templates, adversaries uncover the exact guardrail phrasing and safety boundaries used by enterprise agents, enabling tailored jailbreak payloads and indirect prompt injection attacks.

title: Obot vMCP Unauthorized Prompt and Resource Query
id: a8f4b231-1051-39e1-bc84-obotvmcp0001
status: stable
description: Detects JSON-RPC calls querying prompts or resources via Obot vMCP endpoints originating from untrusted client sessions.
author: Hermes Codex Cyber Intelligence
references:
- https://github.com/obot-platform/obot/security/advisories/GHSA-2c9v-735x-89q5
- https://vulncheck.com/advisories/obot-0.26.0-before-0.26.2-authorization-bypass-via-vmcp-profile-prompts-and-resources
date: 2026-10-10
logsource:
category: application
product: obot
detection:
selection_endpoint:
http_path|contains: '/api/vmcp/'
selection_payload:
request_body|contains:
- '"method":"prompts/list"'
- '"method":"prompts/get"'
- '"method":"resources/list"'
- '"method":"resources/read"'
condition: selection_endpoint and selection_payload
fields:
- user_id
- http_path
- request_body
falsepositives:
- Legitimate authorized administrative agent debugging
level: high

Upgrade to Obot 0.26.2

Immediately upgrade Obot deployments to version 0.26.2 or higher. The patch enforces strict authorization policies across all MCP protocol methods.

Isolate vMCP Connections

Configure distinct MCP client connection pools per tenant rather than relying on global shared component connections.

Audit Exported Prompts

Review all registered MCP prompt templates to verify that no static credentials, API keys, or unmasked PII are embedded in instructions.

Implement Agent Guardrails

Deploy runtime agent monitors to inspect outbound JSON-RPC responses for unauthorized resource leaks and prompt reflections.