Podcast
Hardening Generative AI Chatbots: Prevent Prompt Injection, Data Exfiltration & Hallucinations
Open original DataTalks.Club episode
Hardening Generative AI Chatbots: Prevent Prompt Injection, Data Exfiltration & Hallucinations
Original Episode
Use these links for the canonical episode and media sources.
- Open the original DataTalks.Club podcast page
- Watch on YouTube
- Listen on Spotify
- Listen on Apple Podcasts
Episode Overview
How do you harden generative AI chatbots against prompt injection, data exfiltration, and dangerous hallucinations? In this episode Maria Sukhareva — a principal key expert in AI at Siemens with 15+ years working at the intersection of linguistics and computational AI — walks through real-world risks, attack findings, and practical defenses for chatbot security.
People
Use these links to connect the episode to guest notes.
Chapter Summary
Use these checkpoints to decide whether to open the source transcript.
- 0:00 - Episode Introduction & Guest Overview
- 2:13 - Career Path: From Linguist to Computational Linguistics and Industry
- 4:11 - Role Definition: Principal Key Expert in AI — Advising on Technology and
- 5:42 - Democratization of Generative AI: Rise of Prompting and New “AI Experts”
- 9:28 - Bot Safety Challenge: Large-Scale Chatbot Hacking Exercise and Findings
- 11:38 - Chatbot Failures: Hallucinations, Legal Exposure, and Financial Incidents
- 13:20 - Data Exfiltration Techniques: Overloading Prompts and Knowledge-Base Retrieval
- 16:15 - Mitigations: Output Validation, Query Analysis, and Layered Defenses
- 17:00 - Non-LLM Classifiers: Robust Alternatives to Manipulable Generative Models
- 18:01 - Trust and Hallucinations: User Confidence, Safety, and Adoption Risks
- 20:39 - Chatbot Adoption Issues: Usability, Verbosity, and Return on Investment
- 25:34 - Human-in-the-Loop Solutions: Hybrid Review to Improve Accuracy
- 27:13 - AI as Assistant: Moderation Tools, Autopilot Analogy, and Workforce Impact
- 29:53 - Translation Workflows: AI-Augmented Translators and Quality Control
- 32:28 - Prompt Customization: Controlled Machine Translation with ChatGPT
- 35:44 - Historical Linguistics: Middle & Old English Pronunciation Insights
- 45:08 - Ancient Languages: Cuneiform, Sumerian Transcription, and MT Approaches
- 48:26 - Script Complexity: Logograms vs. Phonetics in Ancient Texts
- 53:01 - Multilingual Models: Progress and Challenges for Low-Resource Languages
- 56:52 - Orthography & Data Quality: Inconsistent Spelling in Historical Corpora
- 57:28 - Industry Trade-offs: Research Innovation vs. ROI and Operational Needs
- 59:14 - Episode Wrap-Up: Key Takeaways on AI Trust, Safety, and Future Directions