Conference Presentation, Keynote
'AI Deployment: Trainwreck or Trailblazer?' Marten Mickos from HackerOne | RAISE Summit 2024 | Paris
- Companies face market pressure to deploy AI now to avoid being perceived as stagnant or losing hiring and consumer appeal, though half of enterprises hesitate due to security concerns.
- Many enterprises risk failure if they ignore safety, with 53% of those deploying AI reporting nothing about security and 17% claiming excellent safety without explanation.
- Specific predicted risks include system breaches, control takeover, data poisoning, prompt injections, brand misrepresentation, hallucinations, toxic content, and operational disruptions.
- Martin Mikos plans to deliver a 15-minute presentation covering these risks to help save money, careers, and prevent embarrassment.
- HackerOne offers a 24/7 network of two million security experts and a proprietary platform named "HAI" that predicts vulnerabilities and automatically generates executive summaries with fixes.
- The HAI platform is designed to evolve daily by training on new data, improving weekly as AI capabilities and adversarial techniques change.
- The HAI platform successfully conducted AI red-teaming tests for Snap to identify offensive image generation vulnerabilities.
- Strategic advice includes staying focused on opportunities while monitoring risks, ensuring deployments are tested by unbiased external teams against adverse scenarios.
- Preparing with proper testing and precautions is predicted to remove exposure, whereas ignoring problems or failing to prepare leads to failure.
- Security and safety are defined as solvable problems encompassing bias, toxic content, and trustworthiness, though adversaries like cybercriminals and nation-states are predicted to increasingly abuse LLMs.
- Martin Mikos asserts that AI will create significantly more wealth and success than problems if deployed correctly, despite valid concerns about both catastrophic failure and salvation.