Automating responses to integrity validation failures

Learn how to use a Cloud Run functions trigger to automatically act on Shielded VM integrity monitoring events.

Overview

Integrity monitoring collects measurements from Shielded VM instances and surfaces them in Cloud Logging. If integrity measurements change across boots of a Shielded VM instance, integrity validation fails. This failure is captured as a logged event, and is also raised in Cloud Monitoring.

Sometimes, Shielded VM integrity measurements change for a legitimate reason. For example, a system update might cause expected changes to the operating system kernel. Because of this, integrity monitoring lets you prompt a Shielded VM instance to learn a new integrity policy baseline in the case of an expected integrity validation failure.

In this tutorial, you'll first create a simple automated system that shuts down Shielded VM instances that fail integrity validation:

Export all integrity monitoring events to a Pub/Sub topic.
Create a Cloud Run functions trigger that uses the events in that topic to identify and shut down Shielded VM instances that fail integrity validation.

Next, you can optionally expand the system so that it prompts Shielded VM instances that fail integrity validation to learn the new baseline if it matches a known good measurement, or to shut down otherwise:

Create a Firestore database to maintain a set of known good integrity baseline measurements.
Update the Cloud Run functions trigger so that it prompts Shielded VM instances that fail integrity validation to learn the new baseline if it is in the database, or else to shut down.

If you choose to implement the expanded solution, use it in the following way:

Each time there is an update that is expected to cause validation failure for a legitimate reason, run that update on a single Shielded VM instance in the instance group.
Using the late boot event from the updated VM instance as a source, add the new policy baseline measurements to the database by creating a new document in the known_good_measurements collection. See Creating a database of known good baseline measurements for more information.
Update the remaining Shielded VM instances. The trigger prompts the remaining instances to learn the new baseline, because it can be verified as known good. See Updating the Cloud Run functions trigger to learn known good baseline for more information.

Prerequisites

Use a project that has Firestore in Native mode selected as the database service. You make this selection when you create the project, and it can't be changed. If your project doesn't use Firestore in Native mode, you will see the message "This project uses another database service" when you open the Firestore console.
Have a Compute Engine Shielded VM instance in that project to serve as the source of integrity baseline measurements. The Shielded VM instance must have been restarted at least once.
Have the gcloud command-line tool installed.
Enable the Cloud Logging and Cloud Run functions APIs by following these steps:
1. In the Google Cloud console, go to the APIs & Services page.
  
  Go to APIs & Services
2. See if Cloud Functions API and Stackdriver Logging API appear on the Enabled APIs and services list.
3. If either of the APIs don't appear, click Add APIs and Services.
4. Search for and enable the APIs, as needed.

Exporting integrity monitoring log entries to a Pub/Sub topic

Use Logging to export all integrity monitoring log entries generated by Shielded VM instances to a Pub/Sub topic. You use this topic as a data source for a Cloud Run functions trigger to automate responses to integrity monitoring events.

Logs Explorer

In the Google Cloud console, go to the Logs Explorer page.

Go to Cloud Logging

In the Query Builder, enter the following values.

resource.type="gce_instance"
AND logName:  "projects/YOUR_PROJECT_ID/logs/compute.googleapis.com/shielded_vm_integrity"

Click Run Filter.
Click More actions, and then select Create sink.
On the Create logs routing sink page:
1. In Sink details, for Sink Name, enter integrity-monitoring, and then click Next.
2. In Sink destination, expand Sink Service, and then select Cloud Pub/Sub.
3. Expand Select a Cloud Pub/Sub topic, and then select Create a topic.
4. In the Create a topic dialog, for Topic ID, enter integrity-monitoring, and then click Create topic.
5. Click Next, and then click Create sink.

Logs Explorer

In the Google Cloud console, go to the Logs Explorer page.

Go to Cloud Logging
Click Options, and then select Go back to Legacy Logs Explorer.
Expand Filter by label or text search, and then click Convert to advanced filter.

Enter the following advanced filter:

resource.type="gce_instance"
AND logName:  "projects/YOUR_PROJECT_ID/logs/compute.googleapis.com/shielded_vm_integrity"

Note that there are two spaces after logName:.

Click Submit Filter.
Click on Create Export.
For Sink Name, enter integrity-monitoring.
For Sink Service, select Cloud Pub/Sub.
Expand Sink Destination, and then click Create new Cloud Pub/Sub topic.
For Name, enter integrity-monitoring and then click Create.
Click Create Sink.

Creating a Cloud Run functions trigger to respond to integrity failures

Create a Cloud Run functions trigger that reads the data in the Pub/Sub topic and that stops any Shielded VM instance that fails integrity validation.

The following code defines the Cloud Run functions trigger. Copy it into a file named main.py.

import base64
import json
import googleapiclient.discovery

def shutdown_vm(data, context):
    """A Cloud Function that shuts down a VM on failed integrity check."""
    log_entry = json.loads(base64.b64decode(data['data']).decode('utf-8'))
    payload = log_entry.get('jsonPayload', {})
    entry_type = payload.get('@type')
    if entry_type != 'type.googleapis.com/cloud_integrity.IntegrityEvent':
      raise TypeError("Unexpected log entry type: %s" % entry_type)

    report_event = (payload.get('earlyBootReportEvent')
        or payload.get('lateBootReportEvent'))

    if report_event is None:
      # We received a different event type, ignore.
      return

    policy_passed = report_event['policyEvaluationPassed']
    if not policy_passed:
      print('Integrity evaluation failed: %s' % report_event)
      print('Shutting down the VM')

      instance_id = log_entry['resource']['labels']['instance_id']
      project_id = log_entry['resource']['labels']['project_id']
      zone = log_entry['resource']['labels']['zone']

      # Shut down the instance.
      compute = googleapiclient.discovery.build(
          'compute', 'v1', cache_discovery=False)

      # Get the instance name from instance id.
      list_result = compute.instances().list(
          project=project_id,
          zone=zone,
              filter='id eq %s' % instance_id).execute()
      if len(list_result['items']) != 1:
        raise KeyError('unexpected number of items: %d'
            % len(list_result['items']))
      instance_name = list_result['items'][0]['name']

      result = compute.instances().stop(project=project_id,
          zone=zone,
          instance=instance_name).execute()
      print('Instance %s in project %s has been scheduled for shut down.'
          % (instance_name, project_id))

In the same location as main.py, create a file named requirements.txt and copy in the following dependencies:
```
google-api-python-client==1.6.6
google-auth==1.4.1
google-auth-httplib2==0.0.3
```
Open a terminal window and navigate to the directory containing main.py and requirements.txt.

Run the gcloud beta functions deploy command to deploy the trigger:

gcloud beta functions deploy shutdown_vm \
    --project PROJECT_ID \
    --runtime python37 \
    --trigger-resource integrity-monitoring \
    --trigger-event google.pubsub.topic.publish

Creating a database of known good baseline measurements

Create a Firestore database to provide a source of known good integrity policy baseline measurements. You must manually add baseline measurements to keep this database up to date.

In the Google Cloud console, go to the VM instances page.

Go to the VM instances page
Click the Shielded VM instance ID to open the VM instance details page.
Under Logs, click on Stackdriver Logging.
Locate the most recent lateBootReportEvent log entry.
Expand the log entry > jsonPayload > lateBootReportEvent > policyMeasurements.
Note the values for the elements contained in lateBootReportEvent > policyMeasurements.
In the Google Cloud console, go to the Firestore page.

Go to the Firestore console
Choose Start collection.
For Collection ID, type known_good_measurements.
For Document ID, type baseline1.
For Field name, type the pcrNum field value from element 0 in lateBootReportEvent > policyMeasurements.
For Field type, select map.
Add three string fields to the map field, named hashAlgo, pcrNum, and value, respectively. Make their values the values of the element 0 fields in lateBootReportEvent > policyMeasurements.
Create more map fields, one for each additional element in lateBootReportEvent > policyMeasurements. Give them the same subfields as the first map field. The values for those subfields should map to those in each of the additional elements.

For example, if you are using a Linux VM, the collection should look similar to the following when you are done:

If you are using a Windows VM, you will see more measurements thus the collection should look similar to the following:

Updating the Cloud Run functions trigger to learn known good baseline

The following code creates a Cloud Run functions trigger that causes any Shielded VM instance that fails integrity validation to learn the new baseline if it is in the database of known good measurements, or else shut down. Copy this code and use it to overwrite the existing code in main.py.

import base64
import json
import googleapiclient.discovery

import firebase_admin
from firebase_admin import credentials
from firebase_admin import firestore

PROJECT_ID = 'PROJECT_ID'

firebase_admin.initialize_app(credentials.ApplicationDefault(), {
    'projectId': PROJECT_ID,
})

def pcr_values_to_dict(pcr_values):
  """Converts a list of PCR values to a dict, keyed by PCR num"""
  result = {}
  for value in pcr_values:
    result[value['pcrNum']] = value
  return result

def instance_id_to_instance_name(compute, zone, project_id, instance_id):
  list_result = compute.instances().list(
      project=project_id,
      zone=zone,
      filter='id eq %s' % instance_id).execute()
  if len(list_result['items']) != 1:
    raise KeyError('unexpected number of items: %d'
        % len(list_result['items']))
  return list_result['items'][0]['name']

def relearn_if_known_good(data, context):
    """A Cloud Function that shuts down a VM on failed integrity check.
    """
    log_entry = json.loads(base64.b64decode(data['data']).decode('utf-8'))
    payload = log_entry.get('jsonPayload', {})
    entry_type = payload.get('@type')
    if entry_type != 'type.googleapis.com/cloud_integrity.IntegrityEvent':
      raise TypeError("Unexpected log entry type: %s" % entry_type)

    # We only send relearn signal upon receiving late boot report event: if
    # early boot measurements are in a known good database, but late boot
    # measurements aren't, and we send relearn signal upon receiving early boot
    # report event, the VM will also relearn late boot policy baseline, which we
    # don't want, because they aren't known good.
    report_event = payload.get('lateBootReportEvent')
    if report_event is None:
      return

    evaluation_passed = report_event['policyEvaluationPassed']
    if evaluation_passed:
      # Policy evaluation passed, nothing to do.
      return

    # See if the new measurement is known good, and if it is, relearn.
    measurements = pcr_values_to_dict(report_event['actualMeasurements'])

    db = firestore.Client()
    kg_ref = db.collection('known_good_measurements')

    # Check current measurements against known good database.
    relearn = False
    for kg in kg_ref.get():

      kg_map = kg.to_dict()

      # Check PCR values for lateBootReportEvent measurements against the known good
      # measurements stored in the Firestore table

      if ('PCR_0' in kg_map and kg_map['PCR_0'] == measurements['PCR_0'] and
          'PCR_4' in kg_map and kg_map['PCR_4'] == measurements['PCR_4'] and
          'PCR_7' in kg_map and kg_map['PCR_7'] == measurements['PCR_7']):

        # Linux VM (3 measurements), only need to check above 3 measurements
        if len(kg_map) == 3:
          relearn = True

        # Windows VM (6 measurements), need to check 3 additional measurements
        elif len(kg_map) == 6:
          if ('PCR_11' in kg_map and kg_map['PCR_11'] == measurements['PCR_11'] and
              'PCR_13' in kg_map and kg_map['PCR_13'] == measurements['PCR_13'] and
              'PCR_14' in kg_map and kg_map['PCR_14'] == measurements['PCR_14']):
            relearn = True

    compute = googleapiclient.discovery.build('compute', 'beta',
        cache_discovery=False)

    instance_id = log_entry['resource']['labels']['instance_id']
    project_id = log_entry['resource']['labels']['project_id']
    zone = log_entry['resource']['labels']['zone']

    instance_name = instance_id_to_instance_name(compute, zone, project_id, instance_id)

    if not relearn:
      # Issue shutdown API call.
      print('New measurement is not known good. Shutting down a VM.')

      result = compute.instances().stop(project=project_id,
          zone=zone,
          instance=instance_name).execute()

      print('Instance %s in project %s has been scheduled for shut down.'
            % (instance_name, project_id))

    else:
      # Issue relearn API call.
      print('New measurement is known good. Relearning...')

      result = compute.instances().setShieldedInstanceIntegrityPolicy(
          project=project_id,
          zone=zone,
          instance=instance_name,
          body={'updateAutoLearnPolicy':True}).execute()

      print('Instance %s in project %s has been scheduled for relearning.'
        % (instance_name, project_id))

Copy the following dependencies and use them to overwrite the existing code in requirements.txt:

google-api-python-client==1.6.6
google-auth==1.4.1
google-auth-httplib2==0.0.3
google-cloud-firestore==0.29.0
firebase-admin==2.13.0

Open a terminal window and navigate to the directory containing main.py and requirements.txt.

Run the gcloud beta functions deploy command to deploy the trigger:

gcloud beta functions deploy relearn_if_known_good \
    --project PROJECT_ID \
    --runtime python37 \
    --trigger-resource integrity-monitoring \
    --trigger-event google.pubsub.topic.publish

Manually delete the previous shutdown_vm function in the cloud function console.
In the Google Cloud console, go to the Cloud Functions page.

Go to Cloud Functions
Select the shutdown_vm function and click delete.

Verify the automated responses to integrity validation failures

First, check if you have a running instance with Secure Boot turned on as a Shielded VM option. If not, you can create a new instance with Shielded VM image (Ubuntu 18.04LTS) and turn on the Secure Boot option. You may be charged a few cents for the instance (this step can be finished within an hour).
Now, assume for some reason, you want to manually upgrade the kernel.
SSH into the instance, and use the following command to check the current kernel.
```
uname -sr
```
You should see something like Linux 4.15.0-1028-gcp.
Download a generic kernel from https://kernel.ubuntu.com/~kernel-ppa/mainline/
Use the command to install.
```
sudo dpkg -i *.deb
```
Reboot the VM.
You should notice the VM is not booting up (cannot SSH into the machine). This is what we expect, because the signature of the new kernel is not in our Secure Boot whitelist. This also demonstrates how Secure Boot can prevent an unauthorized/malicious kernel modification.
But because we know this time the kernel upgrading is not malicious and is indeed done by ourself, we can turn off Secure Boot in order to boot the new kernel.
Shutdown the VM and untick the Secure Boot option, then restart the VM.
The boot of the machine should fail again! But this time it is being shutdown automatically by the cloud function we created as the Secure Boot option has been altered (also because of the new kernel image), and they caused the measurement to be different than the baseline. (We can check that in the cloud function's Stackdriver log.)
Because we know this is not a malicious modification and we know the root cause, we can add the current measurement in lateBootReportEvent to the known good measurement Firebase table. (Remember there are two things being changed: 1. Secure Boot option 2. Kernel Image.)

Follow the previous step Creating a database of known good baseline measurements to append a new baseline to the Firestore database using the actual measurement in the latest lateBootReportEvent.
Now reboot the machine. When you check the Stackdriver log, you will see the lateBootReportEvent still showing false, but the machine should now boot successfully, because the cloud function trusted and relearned the new measurement. We can verify it by checking the Stackdriver of the cloud function.
With Secure Boot being disabled, we can now boot into the kernel. SSH into the machine and check the kernel again, you will see the new kernel version.
```
uname -sr
```
Finally, let's clean up the resources and the data used in this step.
Shutdown the VM if you created one for this step to avoid additional charge.
In the Google Cloud console, go to the VM instances page.

Go to the VM instances page
Remove the known good measurements you added in this step.
In the Google Cloud console, go to the Firestore page.

Go to the Firestore page