Corporations have been slow to adopt generative AI in the past. However, new surveys show that adoption has soared in 2024. We discuss the risks that are holding back companies from deploying AI based systems.

We describe and automated redteaming tool called promptfoo, and interviewed the creator of the project. This tool does blackbox testing using test cases provided by users and grading models.

The Mitre attack framework has been generalized from normal security vulnerabilities to AI specific vulnerabilities with the new Atlas framework. This is another indicator of a maturing industry.

#ai #redteaming #enterprises

Intro | promptfoo
https://www.promptfoo.dev/docs/intro/

Data Suggests Growth in Enterprise Adoption of AI is Due to Widespread Deployment by Early Adopters, But Barriers Keep 40% in the Exploration and Experimentation Phases
https://newsroom.ibm.com/2024-01-10-D...

The state of AI in early 2024: Gen AI adoption spikes and starts to generate value
https://www.mckinsey.com/capabilities...

AI Adoption by Canadian Businesses Slow Despite Stated Interest
https://www.goodmans.ca/insights/post...

The Challenges of Enterprise AI and LLM Adoption - Part 1
  / challenges-enterprise-ai-llm-adoption-part...  

Investors Pour $27.1 Billion Into A.I. Start-Ups, Defying a Downturn
https://www.nytimes.com/2024/07/03/te...

AI Startups Raised $50 Billion Last Year, But Some Investors Are Starting to Pass — Here's Why
https://www.entrepreneur.com/business...

0:00 Intro
0:19 Contents
0:27 Part 1: Corporate use of generative AI
0:37 Three categories of existential risks
1:19 IBM survey of enterprises
1:40 What organizations worry about most
2:14 AI startups raising record amounts
2:57 Develop, fine-tune, or purchase models
3:22 Companies with less AI expertise
3:31 Enterprise AI adoption jumped this year
3:56 Adopting AI in business functions
4:24 1 to 4 months for productionizing
4:48 Details of business functions
5:44 People use generative AI at home and at work
6:18 Example: voice chatbot to nudge users
7:16 How this example helps the business
7:43 Part 2: Tackling risks with automated redteaming
7:55 Path to prototype but not deployment
8:33 Interviewed startup founder Ian Webster
8:48 LLM security vs AI safety
9:08 Promptfoo redteaming tool
9:47 Not a sponsorship
10:04 Redteaming applications rather than frontier models
11:04 Blackbox attack structure
11:35 Promptfoo test cases: LLM graded
12:06 Agentic systems are more complicated
12:22 Greybox and whitebox techniques
13:00 Part 3: Implications of increased adoption
13:27 Increased adoption is occurring worldwide
14:09 Canada not doing so well
14:25 Mitre organization provides CVE security identifiers
15:38 Mitre framework for classifying AI vulnerabilities
16:16 The simple sort of AI
16:46 Compliance standards for AI
17:17 Conclusion
17:44 How Promptfoo fits in
18:32 Outro