Skip to content
Agent Threat Rules (ATR)中文
ΖPrivilege Escalation
ATR-2026-01981
Recorded 2026-07-04 · Author ATR Community

Network-AI ApprovalInbox Unauthenticated Cross-Origin Approval Bypass (GHSA-mxjx-28vx-xjjj)

Detects GHSA-mxjx-28vx-xjjj (CWE-862 Missing Authorization + CWE-352 CSRF via wildcard CORS): network-ai's ApprovalInbox HTTP server (lib/approval-inbox.ts) has no authentication of any kind and sets Access-Control-Allow-Origin: * on every route, including the state-changing POST /approvals/:id/approve and /deny.

unauth approval gate bypass · pattern · confidence: high

Severity
high
Lanes
alert · huntmarks and reports a match
Maturity
testnot yet settled
Scan target
tool_call
Status
experimental

Framework mappings

OWASP Agentic Top 10
  • ASI03:2026
  • ASI06:2026
OWASP LLM Top 10
  • LLM06:2025

ATR wrote these mappings. The framework owners have not reviewed or endorsed them.

Description

Detects GHSA-mxjx-28vx-xjjj (CWE-862 Missing Authorization + CWE-352 CSRF via wildcard CORS): network-ai's ApprovalInbox HTTP server (lib/approval-inbox.ts) has no authentication of any kind and sets Access-Control-Allow-Origin: * on every route, including the state-changing POST /approvals/:id/approve and /deny. ApprovalInbox is the network surface of the human-in-the-loop Approval Gate that gates high-risk operations (writes, shell commands, budget spend). Any caller that can reach the inbox port — a co-located process, an SSRF, a non-loopback remote client, or any website the operator visits in a browser via the wildcard CORS — can enumerate pending approvals (GET /approvals/) and approve them (POST /approvals/:id/approve) with no Authorization header, defeating the Approval Gate so a gated shell command executes without consent. Affected: network-ai <= 5.11.0. Detection covers: (a) a state-changing request to the approval control plane (/approvals/<id>/approve or /deny) with no auth context, matching the unauthenticated-client PoC shape; (b) an enumeration request to the approval queue (GET /approvals/) paired with the wildcard-CORS marker on the approval control plane; (c) explicit GHSA-mxjx-28vx-xjjj / ApprovalInbox exploitation framing. The detection target is the request shape at the approval control plane — an unauthenticated approve/deny or enumerate call combined with the wildcard-CORS marker — which is the exact bypass primitive, caught before the gated action is released.

Suggested responsealertblock input
Author: ATR Community · Created 2026-07-04 · Last modified 2026-11-18

Detection conditions

Combinator: any
  1. 01
    Unauthenticated POST to the approval control plane's approve/deny route — approvedBy defaults to anonymous/attacker with no Authorization header, matching the GHSA-mxjx-28vx-xjjj PoC shape.
    field: contentop: regex
  2. 02
    Enumeration of the pending-approvals queue paired with a wildcard Access-Control-Allow-Origin marker on the approval control plane — the cross-origin approval-inbox reconnaissance step.
    field: contentop: regex
  3. 03
    Explicit GHSA-mxjx-28vx-xjjj reference combined with ApprovalInbox exploitation language — attack framing in a skill or tool description.
    field: contentop: regex

Attacks it catches

  1. tool_description:
    POST /approvals/07d6f277efe35ac1/approve {"approvedBy": "anonymous"}
  2. tool_description:
    GET /approvals/?status=pending -> 200, Access-Control-Allow-Origin: *
  3. tool_description:
    Exploiting GHSA-mxjx-28vx-xjjj: the network-ai ApprovalInbox has no auth, so I can approve any pending action via CORS.

Attack examples the rule must catch, many based on reported attacks and some written to match their shape. They are versioned with the rule as regression tests. If a later revision stops catching one, it shows.

Benign inputs it leaves alone

  1. Authenticated operator approval with a Bearer token present
    tool_description:
    POST /approvals/07d6f277efe35ac1/approve {"approvedBy": "ops-lead", "Authorization": "Bearer sk_live_abc123"}
  2. General question about the feature, not an exploitation request
    tool_description:
    How does the ApprovalInbox approve/deny flow work in network-ai?
  3. Read-only health check on an unrelated route, no wildcard CORS marker present
    tool_description:
    GET /approvals/health
  4. Approve call description explicitly noting mTLS-gated internal deployment, no anonymous/attacker/no-auth marker
    tool_description:
    POST /approvals/07d6f277efe35ac1/approve {"approvedBy": "ops-lead"} behind an internal VPN with mTLS client cert required

Known false positives

  • ▸Legitimate operator approving a pending action through an authenticated ApprovalInbox deployment that has been hardened with a reverse-proxy auth layer — expected to match; review confirms intent.
  • ▸Security advisory or PR text quoting the GHSA-mxjx-28vx-xjjj PoC.
  • ▸A read-only status check against an endpoint that is not the approvals control plane.
title: "Network-AI ApprovalInbox Unauthenticated Cross-Origin Approval Bypass (GHSA-mxjx-28vx-xjjj)"
id: ATR-2026-01981
rule_version: 1
status: experimental
description: >
  Detects GHSA-mxjx-28vx-xjjj (CWE-862 Missing Authorization + CWE-352 CSRF via
  wildcard CORS): network-ai's ApprovalInbox HTTP server (lib/approval-inbox.ts)
  has no authentication of any kind and sets Access-Control-Allow-Origin: * on
  every route, including the state-changing POST /approvals/:id/approve and
  /deny. ApprovalInbox is the network surface of the human-in-the-loop Approval
  Gate that gates high-risk operations (writes, shell commands, budget spend).
  Any caller that can reach the inbox port — a co-located process, an SSRF, a
  non-loopback remote client, or any website the operator visits in a browser
  via the wildcard CORS — can enumerate pending approvals (GET /approvals/) and
  approve them (POST /approvals/:id/approve) with no Authorization header,
  defeating the Approval Gate so a gated shell command executes without
  consent. Affected: network-ai <= 5.11.0.

  Detection covers:
  (a) a state-changing request to the approval control plane
      (/approvals/<id>/approve or /deny) with no auth context, matching the
      unauthenticated-client PoC shape;
  (b) an enumeration request to the approval queue (GET /approvals/) paired
      with the wildcard-CORS marker on the approval control plane;
  (c) explicit GHSA-mxjx-28vx-xjjj / ApprovalInbox exploitation framing.

  The detection target is the request shape at the approval control plane —
  an unauthenticated approve/deny or enumerate call combined with the
  wildcard-CORS marker — which is the exact bypass primitive, caught before
  the gated action is released.
author: "ATR Community"
date: "2026/07/04"
schema_version: "0.1"
detection_tier: pattern
maturity: test
severity: high

references:
  owasp_llm:
    - "LLM06:2025"
Show all (111 more lines)
  owasp_agentic:
    - "ASI03:2026"
    - "ASI06:2026"
  mitre_atlas:
    - "AML.T0049 - Exploit Public-Facing Application"
  mitre_attack:
    - "T1068 - Exploitation for Privilege Escalation"
    - "T1190 - Exploit Public-Facing Application"
  ghsa:
    - "GHSA-mxjx-28vx-xjjj"

metadata_provenance:
  mitre_atlas: human-reviewed
  owasp_llm: human-reviewed
  owasp_agentic: human-reviewed

compliance:
  eu_ai_act:
    - { article: "15", context: "Article 15 cybersecurity — runtime detection of unauthenticated cross-origin approval-gate bypass on network-ai's ApprovalInbox control plane.", strength: primary }
    - { article: "9",  context: "Article 9 risk management — runtime risk-treatment control for unauthenticated cross-origin approval-gate bypass.", strength: secondary }
  nist_ai_rmf:
    - { subcategory: "MP.5.1", context: "MAP 5.1 — adversarial input characterised/detected for unauthenticated cross-origin approval-gate bypass.", strength: primary }
    - { subcategory: "MG.3.2", context: "MANAGE 3.2 — runtime monitoring control for unauthenticated cross-origin approval-gate bypass.", strength: secondary }
  iso_42001:
    - { clause: "8.1", context: "Clause 8.1 operational control — detection of unauthenticated cross-origin approval-gate bypass payload.", strength: primary }
    - { clause: "8.3", context: "Clause 8.3 AI risk treatment — runtime detection as treatment control.", strength: secondary }

tags:
  category: privilege-escalation
  subcategory: unauth-approval-gate-bypass
  scan_target: tool_call
  confidence: high
  source: ghsa-disclosure
  vendor_sources: network-ai-ghsa-mxjx-28vx-xjjj

agent_source:
  type: tool_call
  framework:
    - network-ai
    - any
  provider:
    - any

detection:
  condition: any
  false_positives:
    - "Legitimate operator approving a pending action through an authenticated ApprovalInbox deployment that has been hardened with a reverse-proxy auth layer — expected to match; review confirms intent."
    - "Security advisory or PR text quoting the GHSA-mxjx-28vx-xjjj PoC."
    - "A read-only status check against an endpoint that is not the approvals control plane."
  conditions:
    - field: content
      operator: regex
      value: '(?i)POST\s+/approvals/[a-f0-9]+/(?:approve|deny)\b[\s\S]{0,300}(?:approvedBy["\x27]?\s*[:=]\s*["\x27]?(?:anonymous|attacker)|no\s+auth|Authorization["\x27]?\s*:\s*(?:null|undefined|absent|none))'
      description: >
        Unauthenticated POST to the approval control plane's approve/deny
        route — approvedBy defaults to anonymous/attacker with no
        Authorization header, matching the GHSA-mxjx-28vx-xjjj PoC shape.
    - field: content
      operator: regex
      value: '(?i)GET\s+/approvals/\??(?:status=pending)?\b[\s\S]{0,200}Access-Control-Allow-Origin["\x27]?\s*:\s*["\x27]?\*'
      description: >
        Enumeration of the pending-approvals queue paired with a wildcard
        Access-Control-Allow-Origin marker on the approval control plane —
        the cross-origin approval-inbox reconnaissance step.
    - field: content
      operator: regex
      value: '(?i)GHSA-mxjx-28vx-xjjj[\s\S]{0,200}(?:approval[- ]?inbox|network-ai|approve|CORS)'
      description: >
        Explicit GHSA-mxjx-28vx-xjjj reference combined with ApprovalInbox
        exploitation language — attack framing in a skill or tool
        description.

response:
  actions:
    - alert
    - block_input
  message_template: >
    [ATR-2026-01981] HIGH: network-ai ApprovalInbox unauthenticated
    cross-origin approval bypass detected (GHSA-mxjx-28vx-xjjj, CWE-862 +
    CWE-352). A caller with no Authorization header reached the approval
    control plane's approve/deny or enumeration route, defeating the
    human-in-the-loop Approval Gate. Block the request and upgrade network-ai
    past 5.11.0 / put ApprovalInbox behind an authenticated reverse proxy.

confidence: 78


test_cases:
  true_positives:
    - tool_description: "POST /approvals/07d6f277efe35ac1/approve {\"approvedBy\": \"anonymous\"}"
      expected: triggered
      description: "Unauthenticated approve call defaults approvedBy to anonymous — GHSA-mxjx-28vx-xjjj PoC shape"
    - tool_description: "GET /approvals/?status=pending -> 200, Access-Control-Allow-Origin: *"
      expected: triggered
      description: "Cross-origin enumeration of pending approvals via wildcard CORS marker"
    - tool_description: "Exploiting GHSA-mxjx-28vx-xjjj: the network-ai ApprovalInbox has no auth, so I can approve any pending action via CORS."
      expected: triggered
      description: "Explicit GHSA-mxjx-28vx-xjjj ApprovalInbox exploitation framing"
  true_negatives:
    - tool_description: "POST /approvals/07d6f277efe35ac1/approve {\"approvedBy\": \"ops-lead\", \"Authorization\": \"Bearer sk_live_abc123\"}"
      expected: not_triggered
      description: "Authenticated operator approval with a Bearer token present"
    - tool_description: "How does the ApprovalInbox approve/deny flow work in network-ai?"
      expected: not_triggered
      description: "General question about the feature, not an exploitation request"
    - tool_description: "GET /approvals/health"
      expected: not_triggered
      description: "Read-only health check on an unrelated route, no wildcard CORS marker present"
    - tool_description: "POST /approvals/07d6f277efe35ac1/approve {\"approvedBy\": \"ops-lead\"} behind an internal VPN with mTLS client cert required"
      expected: not_triggered
      description: "Approve call description explicitly noting mTLS-gated internal deployment, no anonymous/attacker/no-auth marker"

Revision history

Created
2026-07-04
Last modified
2026-11-18
Full commit history on GitHub