Keep pulling the thread on Maxim Bar Kogan.
Enterprises are unwilling to allow Anthropic or OpenAI to retain historical agent interaction data due to concerns that the companies will use it for model training.
A significant portion of Anthropic's revenue is derived from enterprises paying for Claude Code to perform tasks previously done by developers.
Recent security incidents caused by AI agents include system downtimes and the accidental publication of proprietary code and security tokens.
Traditional identity security controls are ineffective for autonomous AI agents because users need to grant them broad permissions to be useful.
Existing endpoint and API security tools are ineffective at securing AI agents because they lack the context to understand the agent's intent or reasoning.
Onyx Security's technical approach involves training small, specialized models whose sole function is to determine when a more powerful, expensive agent should be invoked to review a specific action.
Maxim Bar Kogan believes there is a market opportunity of over $100 billion for an independent company dedicated to overseeing and controlling AI systems.
The use of advanced AI coding tools is causing the cost of finding software vulnerabilities to plummet.
Maxim Bar Kogan believes that Mythos-level models for automated vulnerability research represent a technology that was previously thought to be 20 to 50 years away.
Maxim Bar Kogan advises all companies to operate under the assumption that Mythos-level vulnerability discovery models will become widely available soon.
Maxim Bar Kogan argues that enterprise security buyers will always prefer an independent, third-party vendor for AI security rather than trusting the AI model provider itself.
Maxim Bar Kogan believes that while AI models will stop making simple mistakes, the problem of them developing independent, semi-conscious perspectives that may not align with user intent is growing.