CVE-2026-105138: Obot Agent Platform MCP Catalog Entry API Credential Exposure and Lateral Movement
HERMES THREAT SCORE & AGENTIC INFRASTRUCTURE EVALUATION
Target:Obot AI Agent Platform โ MCP Catalog API Handlers & Authorization Layer (pkg/api/handlers/mcp.go) Traditional CVSS v3.1 rates this flaw at 7.7 (High) as an authenticated information disclosure vulnerability within an application boundary. The Hermes Threat Score assesses it at 86 (HIGH) with an Agentic Security Score (HASS) of 94 (CRITICAL). In an autonomous agent architecture, Obot serves as the centralized orchestration orchestrator integrating tools and resources via the Model Context Protocol (MCP). The exposed configurations are not merely generic application settings, but production external credentials configured by enterprise administratorsโsuch as AWS IAM access keys, GitHub personal access tokens, database connection URIs, and enterprise SaaS bearer tokens. A low-privilege employee granted access to an agent workflow can query the MCP catalog endpoint to extract plaintext infrastructure secrets, completely bypassing intended tool sandboxes and achieving lateral breakout into external enterprise environments (MITRE ATLAS AML.T0040 & MITRE ATT&CK T1552).
CVE-2026-105138: Obot Agent Platform MCP Catalog Entry API Credential Exposure and Lateral MovementVULNERABILITY
Software platform affected by security vulnerabilities and agentic attack patterns.
๐ Why is this related? (Evidence & Provenance)
“Confirmed security vulnerability in Obot Autonomous AI Agent Platform documented in Hermes dossier.”
- [vulnerability_report]
- [government_confirmation]CISA verified active exploitation in the wild and mandated federal remediation deadline in KEV entry. — Source: Cybersecurity & Infrastructure Security Agency (CISA): CISA Adds CVE-2026-59822 to Known Exploited Vulnerabilities Catalog (Reliability: VERY_HIGH)
Adversaries search compromise victims for unsecured credentials in files, environment variables, or memory.
๐ Why is this related? (Evidence & Provenance)
“Attack execution telemetry aligns with MITRE ATT&CK technique T1552.”
- [government_confirmation]CISA verified active exploitation in the wild and mandated federal remediation deadline in KEV entry. — Source: Cybersecurity & Infrastructure Security Agency (CISA): CISA Adds CVE-2026-59822 to Known Exploited Vulnerabilities Catalog (Reliability: VERY_HIGH)
1. Technical Context & Affected Architecture Matrix
Section titled โ1. Technical Context & Affected Architecture MatrixโObot coordinates autonomous multi-agent systems, providing a web workspace, user authentication, and tool execution runtimes. To equip agents with capabilities (e.g., querying Jira, writing to Snowflake, pushing to Git repositories), administrators configure MCP servers in the central catalog.
| Parameter | Technical Specification | Operational Details |
|---|---|---|
| CVE Identifier | CVE-2026-105138 | International vulnerability catalog tracking identifier |
| Software Platform | Obot Autonomous AI Agent Platform | Open-source enterprise LLM agent orchestration platform |
| Vulnerable Component | pkg/api/handlers/mcp.go & pkg/api/authz/resources.go | Route handler for GET /api/all-mcps/entries/{entry_id} |
| Vulnerability Class | CWE-522 / CWE-200 | Insufficiently Protected Credentials / Information Disclosure |
| Attack Vector | Network (HTTP REST API) | Authenticated low-privilege HTTP GET request |
| Remediation Release | Obot v0.26.2 | Response field sanitization & credential redaction middleware |
[ Low-Privilege Authenticated User ] | GET /api/all-mcps/entries/{entry_id} with User Token | v +-----------------------------------+ | Obot API Gateway & Router | | pkg/api/handlers/mcp.go | +-----------------------------------+ | Authorization Check: User has READ permission? (YES) Response Sanitization: Redact Secrets? (NO - MISSING!) | v +-----------------------------------+ | Full Catalog Entry Returned | | Cleartext env: | | AWS_SECRET_ACCESS_KEY | | GITHUB_TOKEN | | DB_PASSWORD | +-----------------------------------+ | v +-----------------------------------+ | Adversary Lateral Movement | | Direct Access to Enterprise | | Databases, Cloud & Repos | +-----------------------------------+2. Root Cause Analysis & Vulnerable Code Anatomy
Section titled โ2. Root Cause Analysis & Vulnerable Code AnatomyโThe root cause resides in pkg/api/handlers/mcp.go. When an authenticated user requests details for an MCP catalog entry, the server verifies that the user belongs to the workspace or has read permissions on catalog resources. However, it fails to sanitize the output struct before writing the response:
// Vulnerable handler flow in pkg/api/handlers/mcp.go (prior to v0.26.2)func (s *Server) getMCPEntry(w http.ResponseWriter, r *http.Request) { ctx := r.Context() entryID := chi.URLParam(r, "entry_id")
// Authorization check passes for any authenticated user with catalog visibility entry, err := s.storage.GetMCPEntry(ctx, entryID) if err != nil { http.Error(w, err.Error(), http.StatusNotFound) return }
// FLAW: The raw entry object, including entry.Configuration.Secrets // and entry.Configuration.Env containing static tokens, is serialized directly! w.Header().Set("Content-Type", "application/json") if err := json.NewEncoder(w).Encode(entry); err != nil { http.Error(w, err.Error(), http.StatusInternalServerError) return }}In v0.26.2, the response model was restructured to explicitly sanitize the configuration payload:
// Patched implementation in Obot v0.26.2func (s *Server) getMCPEntry(w http.ResponseWriter, r *http.Request) { // ... // Redact secret key-value pairs before transmitting to non-admin callers sanitizedEntry := entry.SanitizeForUser(ctx, user) json.NewEncoder(w).Encode(sanitizedEntry)}3. Weaponization Chain & Attack Vectors
Section titled โ3. Weaponization Chain & Attack VectorsโAn attacker weaponizes this vulnerability by querying the list of available MCP catalog entries and retrieving the configuration details of each tool:
[Authentication] โโ> [Enumerate MCP Entries] โโ> [Fetch Specific Entry] โโ> [Harvest Cleartext Secrets]Standard User GET /api/all-mcps/entries GET /api/.../{entry_id} Extract AWS/GitHub/DB KeysSession Cookie Catalog List Response Unredacted JSON Body Pivot to External SystemsAttack Request Example
Section titled โAttack Request ExampleโAn attacker logs into the Obot interface with an ordinary team member account and issues the following API call:
GET /api/all-mcps/entries/mcp-ent-98f21bc HTTP/1.1Host: obot.corp.internalAuthorization: Bearer eyJhbGciOiJIUzI1NiIsIn...Accept: application/jsonThe vulnerable server responds with unmasked credentials:
{ "id": "mcp-ent-98f21bc", "name": "corporate-github-mcp", "description": "Internal GitHub Enterprise MCP tool server", "command": "npx", "args": ["-y", "@modelcontextprotocol/server-github"], "env": { "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", "ENTERPRISE_ORG": "cyber-research-dev" }, "created_by": "admin@corp.internal"}With the extracted token, the attacker directly accesses corporate GitHub repositories, bypassing MFA, agent guardrails, and role-based permissions enforced within the Obot web UI.
4. Forensic Triage & Threat Hunting
Section titled โ4. Forensic Triage & Threat HuntingโSecurity teams monitoring Obot deployments should analyze HTTP gateway logs for abnormal bursts of MCP catalog detail queries from non-administrative user accounts.
title: Obot Non-Admin Bulk MCP Catalog Secret Harvestingid: 4e9a712c-9011-4fa3-b187-d4fa23c51138status: experimentaldescription: Detects unusual access frequency to Obot MCP catalog detail endpoints by non-administrative user sessions.author: Hermes Codex Cyber Threat Intelligencedate: 2026-10-08logsource: category: webserver product: obot_api / envoy / ingressdetection: selection: cs-method: 'GET' cs-uri-stem|startswith: '/api/all-mcps/entries/' sc-status: 200 filter_admin: cs-username|contains: - 'admin' - 'system' condition: selection and not filter_adminfields: - c-ip - cs-username - cs-method - cs-uri-stem - sc-statuslevel: hightags: - attack.credential_access - attack.t1552.001 - cve.cve-2026-105138// Azure Sentinel Query: Identify bulk MCP catalog queries from single user identitiesW3CIISLog| where csMethod == "GET" and csUriStem startswith "/api/all-mcps/entries/"| summarize RequestCount = count(), UniqueEntries = dcount(csUriStem) by csUserName, cIP, bin(TimeGenerated, 15m)| where UniqueEntries > 5| order by UniqueEntries descindex=app_obot sourcetype=obot_api uri_path="/api/all-mcps/entries/*" method=GET| stats count, values(uri_path) as queried_entries by user, src_ip| where count > 5| sort - count#!/usr/bin/env bash# Script to verify whether deployed Obot exposes plaintext env secrets on catalog endpointsOBOT_HOST="https://obot.corp.internal"USER_TOKEN="<REPLACE_WITH_TEST_USER_TOKEN>"
echo "[*] Querying Obot MCP catalog entries..."ENTRIES=$(curl -s -H "Authorization: Bearer $USER_TOKEN" "$OBOT_HOST/api/all-mcps/entries" | jq -r '.[].id' 2>/dev/null)
for E in $ENTRIES; do RESP=$(curl -s -H "Authorization: Bearer $USER_TOKEN" "$OBOT_HOST/api/all-mcps/entries/$E") if echo "$RESP" | grep -qiE 'TOKEN|SECRET|PASSWORD|KEY'; then echo "[!] WARNING: Unmasked secrets detected in entry $E" fidone5. Remediation & Hardening Strategy
Section titled โ5. Remediation & Hardening Strategyโ-
Upgrade Obot Instance: Update your Obot deployment to version
0.26.2or later immediately:Terminal window helm upgrade obot oci://ghcr.io/obot-platform/charts/obot --version 0.26.2 -
Rotate All Exposed MCP Credentials: Because all authenticated users previously had read access to configured catalog entries, treat all static secrets, API tokens, and private keys entered into the MCP catalog as compromised. Rotate:
- Cloud provider credentials (AWS IAM keys, GCP service accounts)
- Source code repository tokens (GitHub PATs, GitLab tokens)
- Database connection passwords
- Third-party SaaS API keys (Slack, Salesforce, Jira)
-
Transition to Dynamic Secret Resolution: Rather than storing static credentials in catalog entry configuration blocks, configure Obot to pull secrets from HashiCorp Vault, AWS Secrets Manager, or Kubernetes Secrets at tool execution time.
-
Implement Principle of Least Privilege: Restrict user workspace invitations and ensure only trusted personnel have accounts on clusters orchestrating sensitive enterprise agent tools.