World's Fair 2024
LLM Safeguards: Security Privacy Compliance Anti Hallucination: Daniel Whitenack
Overview
This talk addresses the practical challenges of deploying secure, accurate, and compliant AI systems, particularly focusing on open-access large language models (LLMs) in enterprise environments. It outlines common risks such as hallucinations, supply chain vulnerabilities, data breaches, and prompt injection, and proposes mitigation strategies. The core thesis is that by understanding and proactively addressing these risks with existing and emerging technologies, organizations can build more trustworthy AI applications.
Who should watch
- AI Engineers
- Product Managers
- Builders working with LLMs
- Anyone concerned with AI security, privacy, and compliance
- Teams deploying AI in real-world enterprise settings
Key takeaways
- LLMs can confidently generate inaccurate information, a phenomenon known as hallucination, which poses significant risks in high-stakes applications like medical assistance.
- Supply chain vulnerabilities are a concern, as malicious code could be inserted into model assets or code, especially when relying on open-source packages.
- Model servers, often exposed as APIs, are susceptible to attacks, and prompts may inadvertently contain or leak Personally Identifiable Information (PII) if not properly handled.
- Retrieval-Augmented Generation (RAG) systems, while grounding responses, can introduce data privacy concerns by integrating sensitive company data into prompts.
- Prompt injection attacks aim to trick LLMs into revealing private information or executing unintended actions, especially problematic when LLMs are connected to internal systems.
- Mitigating hallucinations can involve using models fine-tuned for factual consistency between AI output and ground truth data, or employing an "LLM as judge" approach.
- Securing against supply chain risks includes using trusted model registries, verifying model integrity with hashes, and avoiding running untrusted code.
- Protecting against data breaches involves implementing filters to detect and remove PII from prompts, and considering confidential computing technologies like memory encryption or trusted attestation.
Notable quotes
*The text on the internet sometimes represents accuracy and not or it might have outdated information or it could just spew out.*
*There's definitely a challenge here with confident answering of inaccurate information.*
*This is a new way that that enters in so we've got these supply chain vulnerabilities.*
Unofficial community note. Prefer the recording for nuance.