Microsoft Talks AI, the Law and Jailbreaking -- Virtualization Review

Microsoft Talks AI, the Law and Jailbreaking

By David Ramel
06/18/2024

Microsoft's recent buzz about generative AI has had less to do with multimodal models and more to do with the law -- as in improving it and also bending it.

The latter refers to the practice of jailbreaking, a cybersecurity attack technique in which a threat actor bypasses or manipulates AI model safety mechanisms so they do something they shouldn't. That can result in unintended or harmful outputs such as violating policies, executing malicious instructions or producing content that should be filtered out. Essentially, it'a the practice of coercing the AI to generate content it wasn't originally designed for, similar to social engineering attacks on humans.

As with human social engineering, a technique called Crescendo involves multiple interactions with the AI model to eventually manipulate it into producing content that violates its intended use. The attacker can then use the AI model to generate content that is harmful or malicious, such as generating fake news or deepfakes or getting the instructions to make a "Molotov cocktail" firebomb as explained in the recent post, "AI jailbreaks: What they are and how they can be mitigated.

Zero Trust
"Microsoft recommends taking a zero-trust approach towards the implementation of AI; assume that any generative AI model could be susceptible to jailbreaking and limit the potential damage that can be done if it is achieved," Microsoft's guidance says. "This requires a layered approach to mitigate, detect, and respond to jailbreaks."

**[Click on image for larger view.]** Layered Zero Trust *(source: Microsoft).*

'Transforming' and 'Redefining' Legal Support with AI
On the other end of the legal spectrum, Microsoft last week looked at how AI is changing the practice of law in posts titled "Azure OpenAI Service: Transforming legal practices with generative AI solutions" and "All in on AI: Explore Microsoft's journey to redefining legal support with AI."

The latter explains how AI tools like Microsoft Copilot can speed up legal work and enhance its product quality when used by internal teams like Corporate, External, and Legal Affairs (CELA), which is leveraging Copilot and AI to deliver broad impact by prioritizing use cases such as:

Create efficiencies for regulatory work: Quickly summarize regulations, streamline analysis, gather research to stay up to date on industry news and legal developments, draft guidance, and provide actionable insights, empowering legal professionals to stay ahead of the curve.
Strengthen compliance and risk management: Analyze large data sets, help proactively spot possible compliance issues, respond to requests for information, and enable agile and efficient action.
Improve client Interactions: Redirect high volume low-risk inquiries to client self-service capabilities to deliver faster responses.
Enhance advisory services: Quickly find relevant information across sources, including outside counsel work product, to facilitate rapid decision making, and draft communications tailored for different audiences verifying key advisory points are clear and relevant.
Simplify transactions: Condense intricate agreements, pinpoint essential clauses, flag potential risks, compare contracts, compile insights, draft clauses, and research legal structures for increased velocity and better decision-making.
Support pro bono: Impact the world around us by creating efficiencies to enable pro bono volunteers to help more clients in need in less time by using AI to empower volunteers with information they need and automate tasks such as form completion.

On the cloud side of things, meanwhile, Microsoft listed key benefits and applications at work that can help teams throughout the legal field manage contracts more efficiently, reduce risk, ensure compliance and drive better legal business with the help of Azure OpenAI Service:

Enhanced document review: Uses natural language processing to analyze documents, providing relevant insights for legal cases.
Accelerated e-discovery: Quickly identifies, collects, and analyzes large volumes of data from various sources.
Improved efficiency: Reduces the time and resources needed for document review.
Identification of key information: Uncovers critical terms and conditions buried within documents.
Risk management: Assist legal users to consider problematic terms and ensure compliance.
Cognitive translation: Implements AI-driven translation to improve communication across languages.
Accessible contracts: Natural language processing capabilities help users navigate and understand complex legal language.
Enhanced decision-making: Provides insights for more informed strategic decisions.