AI Guardrails & Cybersecurity-AI Agents, Red Teaming HandsOn

Explore embedding models, vector embeddings, and vector search, showing how multidimensional vectors capture semantic meaning and distance-based similarity for tasks like query understanding and clustering.

Explore retrieval augmented generation (rag) and how it optimizes language models by referencing an external knowledge base, with data collection, chunking, embedding, and vector database interactions.

Explore the fundamentals of large language models and their role in conversational AI, education, and content creation, and examine AI guardrails that govern their responsible use.

Explore four major LLM constraints in modern ai applications, including hallucination, bias and ethical concerns, data privacy and security, and output alignment, to guide responsible and innovative deployments.

Explore how hallucination in large language models produces false or irrelevant outputs and learn guardrails to detect, mitigate, and safely deploy AI across high-stakes domains.

Explore bias and ethical concerns in LMS, including fairness, equity, privacy, security, transparency, and accountability, and how guardrails safeguard responsible deployment.

Explore data privacy and data security in the use of LLMs, and learn guardrails that protect personal data, prevent unauthorized disclosure, and preserve trust and reputation.

Learn how output alignment keeps large language models aligned with organizational goals and values, preventing misalignment and guiding responsible use through guardrails for generative artificial intelligence.

Explore how AI guardrails use filters and guidelines to govern data intake. See how they safeguard outputs by blocking inappropriate prompts and correcting responses.

Explore how guardrails enhance reliability, predictability, ethical use, and public trust across e-commerce, mental health, legal, and travel domains, with input and output checks and Guardrails AI and Namo guardrails.

Explore prompts as the intermediary between human intent and generative AI, detailing instruction, context, input data, and output format with examples like bar chart summaries.

Explore how prompts guide generative AI and how prompt injection can bypass guardrails, highlighting instruction, context, input data, and output format. Learn mitigation with prompt guard.

Explore prompt guard, an open source prompt card classifier for llama 3.1 that detects malicious prompts and injected inputs. Fine tune it on your data for precise filtering.

demonstrates running a prompt guard model with 86 million parameters on google colab using hugging face, checking for jailbreak and prompt injection using tokenization, logits, and probability scores.

Explore Llama Guard three, a 3.18 billion parameter pre-trained model fine-tuned for content safety classification that outputs safe or unsafe prompts, supports eight languages, and lists 14 moderation categories.

Learn how Llama guard prompts extend Llama 2 to Llama 3, adding new categories such as information, election, code interpreter, and abuse, with a multilingual prompt format and safety checks.

Explore llama guard three variants, including 1B and 8B text-only models and 11B with vision, and the multimodal prompt formats and safety categories, one image per prompt.

Learn to detect hallucination using faithfulness metrics in a retrieval augmented generation pipeline, and apply open-source models like V3 hallucination judge and the Victory hallucination evaluation model.

Explore how to detect model hallucinations with the phi3-hallucination-judge model using a Colab notebook, tokens, and Pemf transformers pipelines to assess prompts, knowledge base, and responses.

Detect hallucination using a vector-based Hallucination Evaluation model on Hugging Face, comparing premise and hypothesis with retrieved evidence in a rag retrieval process, and review real-world Colab examples.

Learn how Microsoft Presidio acts as a privacy guardrail to identify, assess, and anonymize PII across data types, preventing data leakage and ensuring regulatory compliance for AI pipelines.

Explore Presidio's modular architecture with an analyzer engine that detects PII via pattern matching and NLP recognizers, and an anonymizer engine that masks, replaces, or encrypts identified data.

Learn how the Presidio Analyzer Engine detects PII with pattern matching of credit card numbers, phone numbers, social security numbers, NLP-based recognition, and context analysis, plus extensible custom recognizers.

Detects PII entities in text with a Python based Procedure Analyzer and predefined and custom recognizers, using regex NER and NLP engines to analyze and anonymize data.

Create a custom PII recognizer in Presidio, install packages, import the analyze engine and pattern cognizer, and register a title recognizer with a denial list for text analysis.

Explore the anonymizer engine and its privacy-preserving transformations, including redaction, masking, replacement, and hashing, to protect PII while preserving data utility with configurable entity-type controls.

In this hands-on lecture, learn to configure the anonymizer engine on Presidio Analyzer results, apply replace, redact, mask, hash, or encrypt options, and run a phone-number masking example.

Explore guardrails on AWS Bedrock, including content, sensitive information, and word filters, denied topics, and contextual grounding checks, plus defenses against jailbreaks and prompt injections to prevent hallucinations.

Explore Amazon Bedrock guardrails in a hands-on session, use Bedrock Studio playgrounds for chat, text, and image generation, and manage base and custom models with guardrails.

Create and test guardrails in AWS bedrock, using content filters, denied topics, profanity and PII controls, and grounding checks for an investment firm scenario.

Configure multimodal image guardrails with strong content filters for harmful categories using bedrock's guardrails api to assess prompts and images and block misconduct with high strength.

Garrick scans language models for vulnerabilities, identifying failures and prompt injections. It tests foundation models with probes, generators, and detectors across OpenAI, Hugging Face, and Cohere.

Install Garrick via the CLI, explore probe commands, and run vulnerability tests on models like GPT-2 and ChatGPT 3.5/4.0 from Huggingface.

Explore markdown image exfiltration probes and XSS in Garrick documentation, run tests with GPT 3.5 turbo and GPT 2 on Hugging Face, and review detector results and JSON line reports.

Explore how profanity probes reveal LLM vulnerability to harmful text generation, and how detectors flag terms such as sexual profanity and mental disability during GPT-3.5 turbo testing with YAML configs.