Maxim Bar Kogan, CEO of Onyx Security, says enterprises cannot stop AI agent adoption even as agents publish code and tokens they were not supposed to publish.
Bar Kogan argues enterprises will not let Anthropic or OpenAI share historical agent-behavior data because those labs want to train on it, opening a gap for Onyx.
Bar Kogan says human-in-the-loop review will not scale as agent actions multiply, so Onyx trains small, fast models to decide when to escalate actions to a smarter overseer.
Onyx builds AI agents to police other AI agents
Onyx Security is betting that autonomous agents need an independent AI overseer. Agent actions will grow too fast for human review, and existing security tools cannot see why an agent acts. Small, fast models can watch every action and escalate only critical moments, and enterprises will not hand behavior data to data-hungry model labs.
Takeaways
Autonomous coding agents are the fastest-growing risk. Bar Kogan estimates over 50% of enterprise agents are autonomous coding agents and assistants, about 45% are low-code automations and about 2% are first-party. The autonomous group grows fastest and usually runs without controls. Agents have already published code and tokens by mistake.
Only agents can oversee agents at scale. Bar Kogan says human-in-the-loop review fails when agent actions grow 100x, 1000x or a million times. Onyx trains models and builds agents that oversee other agents, packaged as a secure AI control plane.
Existing security tools miss why an agent acts. Identity systems fall short because agents need broad permissions. Endpoint and API tools lack context, so they miss a coding agent on an unrelated task deleting and recreating a database. A proxy is often not viable because AI runs on cloud infrastructure.
Small fast models decide when to escalate. Having Claude Code watch a copy of itself for every agent fails on cost and latency. Onyx trains small models that escalate only critical actions to a smarter agent, like blitz chess players who calculate deeply only in high-risk moments.
Frontier labs cannot solve agent control alone. Many vendors and open source models will coexist. Labs will fix silly model mistakes, but agents acting on independent judgment need an independent overseer with historical behavior data, and enterprises will not share that with data-hungry Anthropic or OpenAI.
Onyx's bet on autonomy was early but right. AutoGPT gave Bar Kogan an early glimpse of autonomous agents. Security buyers called him too early and he might have run out of money. Reasoning models and autonomous coding agents then arrived and delivered gains low-code platforms did not.
Intense buyer pain outweighs startup size. Onyx has under 100 people, and most research engineers come from an Israeli intelligence unit. Bar Kogan says large companies engage young vendors when pain is strong, and security teams, as revenue preservers, partner with promising early vendors.
What comes next for AI and security
Bar Kogan says the market is not overreacting to Mythos, and that foundational controls such as identity lockdown, firewalls, endpoint detection and AI-specific controls matter most.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
He wants broader access to Mythos-level rollouts and says companies should assume such models are coming, since a Chinese one already exists.
Companies that allow many AI tools will do best, and JPMorgan is selective but adopting quickly.
Security teams will eventually be run by AI agents, so products must serve agents as well as humans.