Scaling responsible AI with guardrails from F5

Tom JohnsonGlobal Product & Privacy Counsel

F5 recently announced two new AI security products, F5 AI Guardrails and F5 AI Red Team. This announcement comes at a pivotal moment when industry regulators are increasingly scrutinizing the use of AI to deliver consumer services.

While there are some AI-specific regulations, such as the well-publicized EU AI Act, regulators in many jurisdictions are using existing regulatory frameworks to ensure that companies are using AI responsibly. Accordingly, even in jurisdictions where there are no AI-specific regulations, enterprises still often have a high burden of compliance to operationalize AI innovations in their technology stacks. Furthermore, many models, including large language models (LLMs), operate as black boxes, which makes using such models more challenging in light of enterprises’ compliance obligations to fairness, transparency, and explainability.

F5 AI Guardrails and AI Red Team, the newest additions to the F5 Application Delivery and Security Platform (ADSP) , add robust compliance controls that help enterprises better understand and correlate AI model inputs and outputs, eliminate biases, and empower human oversight.

All of this can happen in a few clicks, which transforms these insights into active defenses that keep models aligned with an organization’s compliance obligations and internal standards for responsible and ethical AI use. For enterprises holding back their AI deployments due to legal or regulatory risk factors, this is a revolutionary step forward.

Simplify and mitigate AI model risk factors across any use case

F5 AI Guardrails capabilities are instrumental to helping organizations enhance their AI governance strategies, which are critical anytime AI has a direct impact on people’s lives. For example, consider financial institutions that want to use AI to process loan applications. These institutions could find themselves at risk anytime AI denies a loan that a human evaluator would have approved; yet the tricky part is there are very few good tools to understand why a model might deny a loan in the first place. Even with thorough testing of a model prior to use, a lack of explainability won’t satisfy regulatory checks or meet the organization’s own requirements for transparency.

Initially when deploying an AI model, organizations want to vigorously test and validate the model to ensure it won’t go off the rails by divulging sensitive information, reproducing biases, or resulting in unfair decisions for which the company might be legally liable.

F5 AI Red Team offers broad and strategic testing of AI systems: automated adversarial testing of systems at scale to identify both obvious and obscure behaviors and vulnerabilities so teams have a comprehensive view of all risks. These red teaming capabilities are powered by an industry-leading database with over 10,000 new attack prompts added each month. Enterprises can also write their own tests to ensure consistent behaviors as models are updated. This effort helps teams ensure models behave as expected, and inform and adapt their data security, threat management, and governance strategies to outpace emerging risks.

With F5 AI Guardrails, organizations can correlate prompts with responses, set alerts for certain volumes of prompts based on subject matter, and raise critical prompts for evaluation and correction. The solution provides the tools for human oversight and recordkeeping that AI governance teams require.

F5 AI Red Team offers both a 30,000-foot view of why a model is behaving a certain way and a granular view to help explain which features or tokens within a prompt were the most influential in deriving a given inference. Going back to the financial services example, a firm can use AI Red Team functionality to identify what data within a loan application was most impactful to a model’s inference or outcome and what specific data sources it relied on most heavily in its decision making. As a result, organizations will have more data points to map how a model works, identify over- or under-weighted features, and provide feedback for further retraining or fine-tuning. They can also quickly generate documentation to support auditing and compliance efforts.

Remove barriers to AI adoption and accelerate compliance

Simplicity is a key differentiator. This point-and-click solution bypasses the need to create bespoke, in-house tools and complements the rest of F5 ADSP to deliver agile testing, monitoring, and fine-tuning that goes beyond mere alignment with standards. Organizations can accelerate their AI initiatives and deployment of multiple, complex AI models in the face of regulatory headwinds, establishing consumer privacy and fairness as cornerstones of their own service offerings.

It's important to note that even in environments where explicit regulations on the use of AI have yet to be developed, regulators are recontextualizing existing legal frameworks to apply to AI processes. Revisiting financial services yet again, regulators are currently measuring firms’ use of AI against the Fair Credit Reporting Act (FCRA), originally passed in 1970. AI Guardrails helps organizations satisfy their requirements now, while better positioning themselves for new regulations that may materialize in the future.

Fast-track compliant, transparent, risk-averse AI

Following this acquisition and expansion of F5 ADSP capabilities, organizations will have what they need to overcome AI’s black box problem, while demonstrating their risk management posture and data governance to regulators. Some of the most valuable AI use cases that were considered high risk can now be implemented with the appropriate guardrails, human monitoring, and explainability tools. F5 can help organizations get there faster, with a comprehensive platform and simple tools designed for any environment.

For more information, read our previous blog post.

About the Author

Tom JohnsonGlobal Product & Privacy Counsel

More blogs by Tom Johnson