As a part of an ongoing effort to maintain you knowledgeable about our newest work, this weblog publish summarizes some latest publications from the SEI within the areas of counter synthetic intelligence (AI), coordinated vulnerability disclosure for machine studying (ML) and AI, safe improvement, cybersecurity, and synthetic intelligence engineering.
These publications spotlight the newest work from SEI technologists in these areas. This publish features a itemizing of every publication, authors, and hyperlinks the place they are often accessed on the SEI web site.
Counter AI: What Is It and What Can You Do About It?
By Nathan M. VanHoudnos, Carol J. Smith, Matt Churilla, Shing-hon Lau, Lauren McIlvenny, and Greg Touhill
Because the strategic significance of AI will increase, so too does the significance of defending these AI programs. To grasp AI protection, it’s needed to know AI offense—that’s, counter AI. This paper describes counter AI. First, we describe the applied sciences that compose AI programs (the AI Stack) and the way these programs are inbuilt a machine studying operations (MLOps) lifecycle. Second, we describe three sorts of counter-AI assaults throughout the AI Stack and 5 menace fashions detailing when these assaults happen throughout the MLOps lifecycle.
Lastly, based mostly on Software program Engineering Institute analysis and apply in counter AI, we give two suggestions. In the long run, the sphere ought to put money into AI engineering analysis that fosters processes, procedures, and mechanisms that cut back the vulnerabilities and weaknesses being launched into AI programs. Within the close to time period, the sphere ought to develop the processes essential to effectively reply to and mitigate counter-AI assaults, equivalent to constructing an AI Safety Incident Response Workforce and lengthening current cybersecurity processes just like the Laptop Safety Incident Response Workforce Providers Framework.
Learn the SEI white paper.
Classes Discovered in Coordinated Disclosure for Synthetic Intelligence and Machine Studying Methods
by Allen D. Householder, Vijay S. Sarvepalli, Jeff Havrilla, Matt Churilla, Lena Pons, Shing-hon Lau, Nathan M. VanHoudnos, Andrew Kompanek, and Lauren McIlvenny
On this paper, SEI researchers incorporate a number of classes discovered from the coordination of synthetic intelligence (AI) and machine studying (ML) vulnerabilities on the SEI’s CERT Coordination Heart (CERT/CC). In addition they embody their observations of public discussions of AI vulnerability coordination circumstances.
Danger administration throughout the context of AI programs is a quickly evolving and substantial area. Even when restricted to cybersecurity danger administration, AI programs require complete safety, equivalent to what the Nationwide Institute of Requirements and Know-how (NIST) describes in The NIST Cybersecurity Framework (CSF).
On this paper, the authors give attention to one a part of cybersecurity danger administration for AI programs: the CERT/CC’s classes discovered from making use of the Coordinated Vulnerability Disclosure (CVD) course of to reported “vulnerabilities” in AI and ML programs.
Learn the SEI white paper.
On the Design, Growth, and Testing of Trendy APIs
by Alejandro Gomez and Alex Vesey
Utility programming interfaces (APIs) are a basic part of contemporary software program functions; thus, practically all software program engineers are designers or customers of APIs. From meeting instruction labels that present reusable code to the highly effective web-based utility programming interfaces (APIs) of in the present day, APIs allow highly effective abstractions by making the system’s operations obtainable to customers, whereas limiting the small print of how the APIs are applied and thus enhancing flexibility of implementation and facilitating replace.
APIs present entry to difficult performance inside giant codebases labored on by dozens if not lots of of individuals, typically rotating out and in of tasks whereas concurrently coping with altering necessities in an more and more adversarial surroundings. Beneath these circumstances, an API should proceed to behave as anticipated; in any other case, calling functions inherit the unintended habits the API system offers. As programs develop in complexity and dimension, the necessity for clear, concise, and usable APIs will stay.
On this context, this white paper addresses the next questions regarding APIs:
- What’s an API?
- What elements drive API design?
- What qualities do good APIs exhibit?
- What particular socio-technical elements of DevSecOps apply to the event, safety, and operational assist of APIs?
- How are APIs examined, from the programs and software program safety patterns perspective?
- What cybersecurity and different finest practices apply to APIs?
Embracing AI: Unlocking Scalability and Transformation By means of Generative Textual content, Imagery, and Artificial Audio
by Tyler Brooks, Shannon Gallagher, and Dominic A. Ross
The potential of generative synthetic intelligence (AI) extends effectively past automation of current processes, making “digital transformation” a chance for a quickly rising set of functions. On this webcast, Tyler Brooks, Shannon Gallagher, and Dominic Ross intention to demystify AI and illustrate its transformative energy in attaining scalability, adapting to altering landscapes, and driving digital innovation. The audio system discover sensible functions of generative textual content, imagery, and artificial audio, with an emphasis on showcasing how these applied sciences can revolutionize many sorts of workflows.
What attendees will be taught:
- Sensible functions of generative textual content, imagery, and artificial audio
- Influence on the scalability of instructional content material supply
- How artificial audio is remodeling AI training
Evaluating Giant Language Fashions for Cybersecurity Duties: Challenges and Finest Practices
by Jeff Gennari and Samuel J. Perl
How can we successfully use giant language fashions (LLMs) for cybersecurity duties? On this podcast, Jeff Gennari and Sam Perl focus on functions for LLMs in cybersecurity, potential challenges, and proposals for evaluating LLMs.
Hearken to/view the podcast.
Utilizing High quality Attribute Situations for ML Mannequin Take a look at Case Era
by Rachel Brower-Sinning, Grace Lewis, Sebastián Echeverría, and Ipek Ozkaya
Testing of machine studying (ML) fashions is a rising problem for researchers and practitioners alike. Sadly, present apply for testing ML fashions prioritizes testing for mannequin operate and efficiency, whereas typically neglecting the necessities and constraints of the ML-enabled system that integrates the mannequin. This restricted view of testing can result in failures throughout integration, deployment, and operations, contributing to the difficulties of transferring fashions from improvement to manufacturing. This paper presents an strategy based mostly on high quality attribute (QA) situations to elicit and outline system- and model-relevant check circumstances for ML fashions. The QA-based strategy described on this paper has been built-in into MLTE, a course of and power to assist ML mannequin check and analysis. Suggestions from customers of MLTE highlights its effectiveness in testing past mannequin efficiency and figuring out failures early within the improvement course of.
Learn the convention paper.