1 of 30

Trusted AI

Borderless Cyber London

2 of 30

San Francisco

Is Not Dead

Cerebral Valley

3 of 30

Daniel Riedel

Executive Director of GenLab

​

Artist

Futurist

Entrepreneur

​

​

Software, Data & Cyber Best Practices Advocate

Board Member, Treasure Open Standards

​

​

​

​

4 of 30

WILL AI BE AN EXTINCTION EVENT FOR TRUST?

5 of 30

6 of 30

7 of 30

8 of 30

“I’m just a ChatGPT Puppet”

9 of 30

AI’s Role in

  • Abuse
  • Targeting
  • Supremacy
  • Law

​

  • Building Trust
    • Adoption, Transparency, Global Standards & Best Practices

​

10 of 30

Abuse

The new bot net - LLMs for Adversarial Campaigns

11 of 30

AI Concerns

  • Malware
  • Phishing
  • Disinformation
  • Brainstorming and Attack Planning
  • Complex Orchestration

​

  • AI Doesn’t Get Tired, and Will Try Try Again.

12 of 30

Evil GPT Engines`

Monetizing Criminal Activity Through Generative AI

13 of 30

Rogue LLMs

Availability - Open Source LLM models are Easy to Startup and Deploy

​

Machiavellian - Rogue LLMs will have no rules, ethics or morality built into their operations

​

Economics - Rogue LLMs will be as powerful, as the money applied, and the money to be gained.

14 of 30

Target

Expanding Attack Surface

15 of 30

Data Poisoning

  • Poisoning the LLM
    • Directly Attacking the LLM
  • Poisoning the DataSet
    • Attacking a known data set
  • Poisoning the Web
    • General global data poisoning to corrupt ingest

Google Spam Filter

Microsoft Twitter ChatBot Tay

16 of 30

Prompt Injection

  • Makes SQL Injection like a walk in the park
  • The challenges with non deterministic systems
  • Building defenses around them
  • Jailbreaking

​

​

Even though you can’t fully protect against prompt injection attacks, I’ll suggest a high-level approach developers can use to consider the risks and reduce their exposure to these attacks.

​

Grandma Attack - Tricking an AI in breaking a guardrail can expose other data, so enterprises and governments have to think carefully about data ingest.

We are building systems that are impossible to fully protect, probabilistic defense.

17 of 30

Supremacy

driving Balkanization & Supply Chain Conflict

18 of 30

Suspicion, IP Theft, Borders

https://llmshield.com/llm-security

19 of 30

Energy, Chips, Software

Supply Chain Attacks

Physical

Network

Software

Data

20 of 30

Culture

  • Balkanization of AI that can harm trust.
  • Values are regional, one LLM is not sustainable.
  • The complexity of values is exponential to the consumption of data.
  • Building guardrails per region will be complex.
  • Regional disinformation as a weapon on itself

​

21 of 30

Law

Legal Challenges, Copyright, Privacy, Data Governance

22 of 30

Legal cases are stacking up

September 5, 2023, T. et al v. OpenAI LP et al

July 14, 2023, Walters v. OpenAI LLC

July 7, 2023, Silverman, et al v. OpenAI Inc.

June 28, 2023, Tremblay v. OpenAI Inc.

June 28, 2023, PM et al v. OpenAI LP et al

23 of 30

How to Build Trust

Transparency, Global Standards & Best Practices

24 of 30

Adoption is critical

Fear of Generative AI is counter productive

Adoption and learning, utilizing the tools to accelerate your own personal workflow

Move Faster - Technical Education - Economic Vitality

​

Global Education, Generative AI has already shown abilities to accelerate learning

​

25 of 30

Trusted AI/Responsible AI Model

Fair/Impartial

Robust/Reliable

Transparent/Explainable

​

Responsible/Accountable

​

Safe/Secure

Privacy/Attribution

Concepts are Still Very New

Lack of Global Consensus

A Continued Conversation

26 of 30

Signing AI & Data

Attribution

Provenance

Intellectual Property

Explainability

Copyright

Safety

27 of 30

Public / Private Partnerships

Best Practices

Frameworks

Cyber Security & Disinformation Standards and Norms

Data Sharing Standards (Sign that data)

HippoCratic Oath of Development

​

​

A lot of work to do

28 of 30

Standards & Governance is crucial

Cross Border Cooperation is Key!

29 of 30

Thank you

Questions?

30 of 30

References