Maxim Bar Kogan is co-founder and CEO of Onyx Security. He previously worked in offensive cyber and AI. At Onyx he focuses on how enterprises can set and enforce limits on autonomous agents. — Introducing the Secure AI Control Plane.

Visual summary of operating lessons from Maxim Bar Kogan.

Part 1: The Transition to Agentic Operations

  1. Plan for a rise in delegated actions: In his July 2026 funding announcement, Bar Kogan predicted that agents would soon carry out much more enterprise work. He used that forecast to argue for building oversight before adoption expands. — Onyx: Keeping Humans in Control.
  2. Ask how people will retain control as they delegate decisions: Bar Kogan says he started Onyx with a question about how people can retain control when they rely on systems whose decisions they may struggle to understand. — Introducing the Secure AI Control Plane.
  3. Distinguish generating an answer from carrying out a task: Bar Kogan recalls AutoGPT as a turning point for him. It helped him imagine what models could do when they could use tools and continue taking actions toward a goal. — No Priors: Building an AI Guardian.
  4. Build oversight into critical deployments: Bar Kogan argues that organizations should put safeguards in place before delegating more work to agents in critical industries. — Onyx: Keeping Humans in Control.
  5. Keep core security checks deterministic: Bar Kogan sees value in flexible AI workflows while arguing that controls such as authentication, verification and encryption need predictable behavior. — Inflection Point: Deterministic Controls.
  6. Treat agents as actors that work across systems: Bar Kogan points out that an agent can write code, triage an IT ticket or manage a customer task while moving across tools and identity boundaries. That makes the agent’s activity relevant across several applications. — Introducing the Secure AI Control Plane.
  7. Enforce policy where an agent takes action: Bar Kogan describes oversight across browsers, endpoints, SaaS and cloud systems. He wants controls to check actions before those actions change downstream systems. — Onyx: Keeping Humans in Control.

Part 2: The Inadequacy of Traditional Security

  1. Account for changing context and untrusted inputs: Bar Kogan says an agent’s actions can depend on its model, the context it sees and the outputs of its tools. A policy that assumes the same action sequence each time can miss that variation. — Introducing the Secure AI Control Plane.
  2. Reserve human review for actions that need it: Bar Kogan argues that a person cannot review every action once agents start taking many more actions. Onyx’s proposed workflow uses automated checks and can ask a person to approve a specific action. — No Priors: Building an AI Guardian.
  3. Evaluate an action in the context of the task: Bar Kogan gives the example of deleting a database. That can be part of an approved request to rebuild it, or an unwanted side effect of unrelated work. He argues that endpoint and API tools can lack the task context needed to tell the difference. — No Priors: Building an AI Guardian.
  4. Prepare defenses before more capable models become widely available: Bar Kogan argues that a staged model release can give defenders time to build playbooks and share what they learn. A comparable release elsewhere could shorten that preparation window. — No Priors: Building an AI Guardian.
  5. Look for new security problems as the underlying technology changes: Bar Kogan argues that the move to AI gives smaller security companies a chance to serve needs incumbents have yet to address. — Calcalist Interview on Onyx’s Series B.
  6. Define what the agent is allowed to do: Bar Kogan says governance should specify the actions that are within an agent’s authority, in addition to detecting security threats. — Calcalist Interview on Onyx’s Launch.

Part 3: The Architecture of AI Oversight

  1. Use a separate system to assess agent activity: Bar Kogan describes Onyx’s approach as training models to assess actions proposed by other agents, so the system can decide which actions merit intervention. — No Priors: Building an AI Guardian.
  2. Spend more computation where the risk is higher: Bar Kogan compares oversight to chess. Players rely on practiced judgment for many moves and spend longer evaluating the moves where getting it wrong would matter most. — No Priors: Building an AI Guardian.
  3. Train small models to decide when to escalate: Bar Kogan describes small models that are specialized for deciding whether to send an action to a more capable model. He frames the choice as a tradeoff among missed risks, latency and cost. — No Priors: Building an AI Guardian.
  4. Intervene before an unsafe action takes effect: Bar Kogan says Onyx’s Guardian Agent can block an action, ask a person for approval, narrow its scope or redirect the agent before the action reaches another system. — Introducing the Secure AI Control Plane.
  5. Bring security and AI expertise into the same team: Bar Kogan credits Onyx’s mix of offensive security and AI research to the founders’ different backgrounds. He says oversight requires expertise in both security and how models behave. — Introducing the Secure AI Control Plane.
  6. Inventory agents and the systems they depend on: Bar Kogan describes discovering models, AI applications and tools connected through MCP, alongside the agents themselves, to understand where runtime policies need to apply. — Introducing the Secure AI Control Plane.
  7. Track oversight at production scale: In its Series B announcement, Onyx reported that it secured more than 1.1 million agents and had inspected more than 66 million AI sessions. Those are company-reported figures. — Onyx: Keeping Humans in Control.

Part 4: Alignment and Independent Governance

  1. Watch for actions that diverge from the user’s goal: Bar Kogan distinguishes simple errors from cases where a model selects an action that may be coherent on its own but differs from what the user wants. He expects the latter to remain important as models improve. — No Priors: Building an AI Guardian.
  2. Use oversight that can assess the model independently: Bar Kogan argues that an enterprise can benefit from an independent assessor whose credibility depends on identifying risks correctly, even if the model vendor also runs its own safety checks. — No Priors: Building an AI Guardian.
  3. Balance controlled access against defenders’ need for tools: Bar Kogan says limiting access to a capable model can reduce immediate exposure. He also argues for giving more defenders access while they still have time to learn how to use it. — No Priors: Building an AI Guardian.
  4. Set explicit limits on access to historical agent data: Bar Kogan says some enterprises are reluctant to share historical agent activity with model vendors because they are concerned about how the data could be used. — No Priors: Building an AI Guardian.
  5. Separate investor conviction from market outcomes: Bar Kogan interprets the size of Onyx’s Series B as evidence that investors believe AI security could become a major cybersecurity category. — Calcalist Interview on Onyx’s Series B.

Part 5: Navigating the Competitive Landscape

  1. Prioritize the customer’s problem in a crowded market: Bar Kogan says Onyx tries to stay focused on what its customers need. He sees listening to them and solving their problems as a better guide for product work than tracking competitors. — Inflection Point: Focus on Customers.
  2. Recognize when a startup depends on a market arriving in time: Bar Kogan recalls security colleagues asking whether enterprises would adopt agents before Onyx ran out of money. He acknowledges that the timing could have gone against the company. — No Priors: Building an AI Guardian.
  3. Use funding to build the controls a growing market will need: Bar Kogan announced a $113 million Series B and framed the round as an investment in the platform needed to govern more agent activity. — Onyx: Keeping Humans in Control.
  4. Build toward independence if that is the goal: Bar Kogan says his goal for Onyx is to build an independent global company. He describes that ambition in the context of investors looking for category leaders. — Calcalist Interview on Onyx’s Series B.
  5. Connect oversight to visible business results: Bar Kogan points to software development and customer interactions as two areas where he sees measurable returns from AI deployment. He argues that defining what agents may do helps leaders understand those returns. — Calcalist Interview on Onyx’s Launch.