No PriorsBuilding an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan
CHAPTERS
AI agents are scaling fast—and so are “bad actions”
Maxim frames the core risk: as enterprises deploy more autonomous agents, the number of actions explodes and so does the probability of harmful or unintended behavior. Recent examples include agents leaking secrets, publishing tokens, or causing downtime—issues enterprises can’t solve by simply slowing adoption.
Why Onyx bet early on autonomous agents (AutoGPT as the catalyst)
Sarah and Maxim revisit the company’s origin thesis: security should focus on what agents do, not just what employees type into chatbots. AutoGPT provided an early glimpse of tool-using autonomy, and Maxim’s “AGI-pilled” view pushed him to prioritize oversight of increasingly capable systems.
What Onyx builds: a guardian layer and a secure AI control plane
Maxim gives a crisp product breakdown: Onyx trains models/agents to oversee other agents and productizes this into a control plane that discovers and hooks into enterprise AI systems. The purpose is to verify legitimacy of actions at machine scale when humans can’t keep up.
Enterprise agent adoption today: three buckets and what’s growing fastest
Onyx categorizes deployments into low-code automation platforms, first-party internal agents, and highly autonomous coding agents. Maxim observes that autonomous coding assistants are already a majority share and are growing the fastest—often with minimal controls.
Why traditional security controls struggle with agentic systems
Existing controls (identity, endpoint, API security) lack the contextual understanding of an agent’s intent and plan. Because enterprises often want agents to act with broad permissions, classic least-privilege approaches become harder to apply without destroying usefulness.
Why “just use a proxy + policy engine” isn’t sufficient
Maxim distinguishes integration from solving the core problem: even if you can proxy traffic, the hard part is deciding whether an action is appropriate. Many agent environments also can’t be neatly proxied due to where they run (cloud, third-party infra, endpoints).
Why Onyx trains its own models: small specialists that call in “smart review”
Onyx’s approach is to avoid running expensive frontier-model oversight for every action. Instead, they train small, specialized models that decide when to escalate to deeper inspection—optimizing for cost, latency, and reliability while still catching risky behavior.
The blitz-chess analogy: intuition most of the time, deep calculation at critical moments
Sarah proposes—and Maxim agrees—that effective oversight resembles high-level blitz chess: fast intuitive moves most of the time, with occasional slow, intensive calculation in critical positions. This mirrors selective allocation of compute to moments of high risk.
Onyx’s talent DNA and Israel’s evolving AI-security ecosystem
Maxim describes a hybrid culture combining cyber experience with deep AI research, reflecting both founders’ backgrounds. He argues Israel is catching up quickly in AI (models, infra, chips) while retaining strong security product instincts rooted in close contact with security buyers.
Mechanistic interpretability as part of the long-term control solution
Maxim defends the bet that understanding internal model structure (weights/activations) will matter for governance and safety. He suggests smarter-than-human models may help crack interpretability, enabling better understanding of intelligence and model behavior.
Earning trust with Fortune-scale customers as a young company
Sarah probes the trust gap: why would major enterprises rely on a small startup? Maxim argues acute pain and urgent risk drive buyers to evaluate new vendors, and security leaders prefer early partnership over doing nothing while adoption accelerates.
Mythos and the collapse in cost of vulnerability discovery—how to respond
They discuss the security shock from AI-assisted vulnerability research (Mythos as shorthand). Maxim argues the market isn’t overreacting: teams need quick mitigations now, but must build foundational controls across the stack—and AI needs its own foundational security layer.
Controlled release of powerful security-relevant models (Glasswing/Daybreak)
Maxim weighs gradual rollout: it buys defenders time, but could be disastrous if adversaries gain equivalent capability first. His recommendation is to assume these models will arrive regardless and to prepare with foundational controls now.
Enterprise holdouts, tool diversity, and Onyx’s strategy in a fast-moving landscape
Bans are rarer now; some regulated firms limit tools but still adopt agents. Maxim argues enterprises should allow multiple tools because leadership in models shifts quickly, while Onyx stays focused on stable primitives (LLMs + tool-using agents) while holding flexible views about post-2026 paradigms.
Why model labs won’t fully own trust/governance—and why independence matters
Maxim argues buyers want an independent certifier rather than the same vendor selling the model. He also notes enterprises won’t share rich historical behavior data with labs due to training concerns, and the multi-model world (closed + open) makes uniform lab-provided security unrealistic.
What the broader tech world misses about security teams—and an AGI-shaped future
Maxim highlights Israel’s strength in understanding how security organizations actually operate day-to-day, which is essential for product adoption. He also reconciles being “AGI-pilled” with building for today’s human buyers while anticipating security workflows increasingly executed by agents.
