Defense at Machine Speed: The Emerging Architecture Powering AI-Native Cybersecurity

Menlo Ventures, the VC firm behind bets like Abnormal AI, Obsidian, and Zafran, just laid out a thesis that's less speculative and more urgent warning: the old model of security teams manually chasing alerts is functionally dead. Attackers, the firm argues, have already crossed into machine-speed operations, running thousands of personalized phishing campaigns from a single operator and pulling off the first documented AI-orchestrated nation-state intrusions. And the target has shifted too. Most breaches now aren't software bugs getting exploited, they're trusted identities and credentials being abused, which means there's nothing to patch. You can't fix a vulnerability that doesn't exist.

The scale problem gets worse fast. Menlo points out that a typical SOC fields thousands of alerts daily but can genuinely investigate maybe a few dozen. Everything else slides by unexamined, and attackers are counting on it. Add tens of thousands of autonomous agents per enterprise, each transacting and messaging on someone's behalf, generating behavior nobody explicitly coded, and the attack surface stops being your network and becomes every identity, human or machine, acting inside your business.

Menlo's answer is a three-layer stack. First, a behavioral context engine that learns what normal looks like for every identity by digging into email, calendar, chat, and identity systems, then flags departures from that baseline. Abnormal AI, a Menlo portfolio company already protecting over a quarter of the Fortune 500, built this by starting in email, tracking who talks to whom and how money actually moves, so it catches attacks that carry no malware and trip no rule-based alarm. Second, autonomous agents that act on findings without waiting in a human queue, exemplified by Zafran's approach to vulnerability remediation, which figures out what's actually exploitable given existing controls and fixes it automatically instead of letting it sit for weeks. Third, continuous validation, split between Obsidian checking internal posture drift and Armadin running swarms of AI agents that behave like sophisticated attackers to pressure-test defenses nonstop rather than during an annual pentest.

What makes this bigger than a typical security pitch is the economic framing. Menlo estimates enterprises will spend trillions over the next decade deploying AI agents as digital labor, and a bank running millions of agents, not thousands, isn't a hypothetical. No board signs off on that without some way to guarantee the agents behave, which is why Cequence, another portfolio company, focuses on governing what an agent can actually do down to individual tool calls. The firm's broader argument is that enterprises won't trust model providers to police their own agents; they'll want an independent layer they control, the same instinct that's driven third-party auditing for decades.

Menlo frames this as a closing window, borrowing language from Abnormal's CEO Evan Reiser about compounding advantages and disadvantages reaching a scale civilization hasn't seen before. Whether that urgency is fully earned or partly a pitch deck for the firm's own portfolio is worth sitting with, but the underlying diagnosis, that identity-based attacks and agent sprawl are outpacing manual defense, tracks with what plenty of security engineers have been saying independently.