Google’s secure AI framework

Previously, you learned that a risk management framework is a set of practices, processes,
and technologies that help enable an organization to identify, assess, analyze, and manage risk
within an organization. As artificial intelligence (AI) becomes more widely adopted, it’s
important for organizations to establish a risk management framework that addresses AI’s
complexities.

In this reading, you’ll learn about Google’s Secure AI Framework (SAIF), and explore examples
of how this framework can be used as a tool to help secure AI technologies and processes.

Google’s Secure AI Framework (SAIF)

The SAIF provides organizations with considerations to make when implementing AI systems
and identifying how to mitigate risks. The SAIF offers six core elements:
1. Expand strong security foundations to the AI ecosystem
2. Extend detection and response to bring AI into an organization’s threat universe
3. Automate defenses to keep pace with existing and new threats
4. Harmonize platform level controls to ensure consistent security across the organization
5. Adapt controls to adjust mitigations and create faster feedback loops for AI
deployment
6. Contextualize AI system risks in surrounding business processes

SAIF core elements

The SAIF is a set of AI-focused best practices that security professionals can implement
throughout the development lifecycles, from infrastructure to data. Organizations should tailor
these best practices to meet their own unique business needs. Also, organizations should form
teams with a variety of stakeholders to guide AI implementation, including security analysts,
cloud engineers, developers, and responsible AI, ethics, and legal team members.

Expand strong security foundations to the AI ecosystem

Organizations should assess the existing security controls in place that secure their underlying
infrastructure, data, and applications. Many, if not all of these controls also apply to securing AI
systems, but some may need to be reconfigured to address the AI system’s specific
requirements. For example, data encryption might traditionally be used for authorization
purposes, but when considering AI, organizations will also need encryption to help prevent
data tampering or theft.

1
Extend detection and response to bring AI into an organization’s threat universe

Organizations must be able to detect and respond to AI-related cyber threats and attacks. One
way organizations accomplish this is by using threat intelligence. Threat intelligence is
evidence-based threat information that provides context about existing or emerging threats.
Imagine a system that’s always up to date with how a virus, attack, or deception might
compromise it. Then, that system can actively alert and initiate a response from analysts, all by
using the power of AI-assisted threat intelligence.

To prepare for threats or cyber attacks, organizations should be familiar with training data used
for AI systems, and the different components of the system. This means the organization
should know how current their system is for training the data, and how its models work with
other security toolsets. With this information, organizations can help identify vulnerabilities
within their AI systems, which contributes to their threat intelligence, and helps inform a
response to potential threats.

It’s also important to monitor the input and output created by generative AI (GenAI) systems.
For example, because of GenAI’s ability to create a vast amount of content, the output should
be closely monitored for negative or harmful output that can be generated for malicious
purposes. Content safety policies can help prevent this risk of harmful content. This content
might appear as GenAI-created malicious emails or misleading output from a company’s
AI-powered chatbot.

Automate defenses to keep pace with existing and new threats

Although AI technology may introduce new threats in the cyber space, AI can also be used as a
tool to defend against those threats. AI innovates recognizing new patterns and attacks sooner
than current toolsets. AI can help detect bias, security breaches, and malicious content. But,
human input is necessary for larger and more important decision-making factors, like
determining what constitutes a breach, and the appropriate response.

Harmonize platform level controls to ensure consistent security across the

organization

AI should be regularly monitored to ensure the AI system, including training data and any
related applications, are performing as expected. A comprehensive understanding of each
component in the AI system’s lifecycle provides organizations with a clear picture of the assets
involved. With this information, the organization can identify any overlapping areas where
controls or frameworks are duplicating efforts. Then, organizations can consolidate
frameworks and controls, helping them save costs and become more efficient.

2
Adapt controls to adjust mitigations and create faster feedback loops for AI
deployment

Threat actors execute new cyberattacks every day, and AI opens a new environment for these
actors to exploit. To help combat evolving threats, organizations should constantly test their AI
implementations to improve the security of the changing threat landscape. One way to test
the durability of an AI system is to perform red team exercises. A red team is a group of
ethical hackers who mimic potential adversaries in order to examine the security defenses of
an organization. Red teams can be a team within an organization, or they can be hired as a
third party.

Organizations, and especially security analysts, should stay up to date on the latest attack
methods and techniques that threat actors use. Common AI vulnerabilities might include
leaking data or forming wrong predictions. Organizations can use red teams or other testing
measures to help identify attack vectors that could exploit these vulnerabilities.

Communication is a critical aspect for AI deployment. Any vulnerability should be shared with
remediation teams, and also with the organization at large. With constant testing in the AI
environment, there should also be constant feedback to stakeholders and to the training data
used for the AI model.

Contextualize AI system risks in surrounding business processes

Organizations should create processes for managing risks related to AI systems. Risk
assessments outline how the organization, third parties, and security controls interact with the
AI model’s components. The risk assessments should encompass the entire AI lifecycle,
including training data and applications that use that data. For example, organizations can
incorporate automated security checks as a control to monitor AI performance both internally,
and with third-party applications. Understanding how and where AI is implemented can help
analysts select the right risks to address first, from inadequately tuned chatbots, to
infrastructure as code repeating AI-generated design mistakes.

Key takeaways
Risk management frameworks are an important tool to help reduce risks and improve security
posture. With the emergence of AI technology, organizations need to integrate frameworks
that address risks AI can impose. Google’s SAIF is a tool that can help organizations navigate
implementing AI and the many considerations involved in the process.

Resources for more information

For more information about securing AI frameworks, check out:
● Google’s quick guide to implementing the SAIF