Video · GitHub ·
Niels Provos on creating bcrypt, defending Google, and securing AI agents
In short
Niels Provos explains how durable security tools such as bcrypt, Google Safe Browsing, and HoneyD emerged from adversarial thinking, and why AI agents now require stronger controls against prompt injection and intent drift. The discussion matters to AI builders because current models can uncover long-hidden vulnerabilities, while deterministic policies, human review, and proof-of-exploit workflows are needed to deploy agents and maintain open-source software safely.
Key points
- bcrypt has remained in use for about 28 years because its iteration cost can be increased as computing power grows.
- Anthropic’s model found a 27-year-old OpenBSD vulnerability, showing that AI can uncover bugs missed for decades.
- Project Iron Curtain converts plain-English rules into deterministic policies for controlling AI agents, but still requires human approval.
- HoneyD’s virtual honeypots inspired agent honeypots designed to expose prompt injection and unauthorized data requests.
- Google Safe Browsing protected more than 4 billion browsers by identifying malicious websites with large-scale machine learning.
- AI-generated vulnerability reports need proof-of-concept exploits so maintainers can distinguish real issues from false positives.
Written by an AI model from the video’s captions and description.
Sit down with legendary cybersecurity researcher Niels Provos as he reflects on three decades of breaking, building, and securing software.



