AI-Native Cybersecurity: Agents Attacking & Defending
1. Thesis
Cybersecurity is shifting from tools that alert humans to agents that act autonomously on both sides — and value is migrating from detection (now commoditizing) to the orchestration and validation layer that decides what machines are allowed to do at machine speed.
---
2. Segments
A. Autonomous SOC / AI Security Analysts
Agents that triage, investigate, and remediate alerts, replacing Tier-1/2 analyst labor.
- Dropzone AI – autonomous alert investigation
- Prophet Security – agentic SOC analyst
- Radiant Security – AI SOC co-pilot/automation
- Torq (HyperSOC) – agentic SOAR evolution
- Microsoft Security Copilot, CrowdStrike Charlotte AI – incumbent embedded agents
Dynamics: Hottest funding category; economics are compelling (analyst labor is ~40% of SOC cost), but incumbents with data gravity (CrowdStrike, Microsoft, Palo Alto's XSIAM) can bundle it, squeezing standalone players toward acquisition.
B. Offensive AI / Autonomous Pentesting & Validation
Agents that continuously attack your own environment to find exploitable paths.
- Horizon3.ai (NodeZero) – autonomous pentesting
- XBOW – AI-driven vulnerability discovery (topped HackerOne leaderboards)
- Pentera – automated security validation
- Picus Security, SafeBreach – breach & attack simulation
- RunSybil (less sure on traction; founded by ex-OpenAI security researcher) – AI pentesting
Dynamics: Structural tailwind — attackers are already using LLMs, so continuous validation replaces annual pentests; a services-to-software conversion of a ~$2B+ pentest market with much better margins.
C. Securing AI Itself (LLM/Agent Security, Guardrails, Red-teaming)
Protecting models, prompts, and agent actions from injection, jailbreaks, and data leakage.
- Protect AI (acquired by Palo Alto Networks, 2025)
- HiddenLayer – model security/detection
- Lakera – prompt injection & guardrails (acquired by Check Point — reasonably confident, announced 2025)
- Prompt Security (acquired by SentinelOne — less sure on close status)
- Robust Intelligence (acquired by Cisco)
Dynamics: Fast consolidation — incumbents are buying rather than building, meaning standalones face a short window; likely a feature layer, not a durable category.
D. Agent Identity & Non-Human Identity (NHI) Security
Governing credentials, permissions, and trust for the exploding population of machine/agent identities.
- Astrix Security
- Oasis Security
- Aembit – workload IAM
- Clutch Security
- Okta / SailPoint – incumbents extending into NHI
Dynamics: Agents outnumber humans 10–50x in identities; earliest-stage segment but the control point for everything agents do — classic identity-layer economics (sticky, horizontal).
E. AI-Powered Email/Social Engineering Defense
Countering AI-generated phishing, deepfakes, and BEC.
- Abnormal AI – behavioral email security
- Material Security
- Reality Defender – deepfake detection
- Sublime Security
Dynamics: Attacker adoption of LLMs made legacy secure email gateways obsolete first; Abnormal proved the AI-native replacement playbook at scale, but the segment is maturing.
---
3. Where Value Concentrates
Segment A (Autonomous SOC) captures the most absolute value — it directly substitutes the largest cost line in security (human analyst labor, a $100B+ services/labor pool) rather than a tooling budget, and it becomes the control plane through which all other detections flow. But the best risk-adjusted venture returns are likely in Segment D (agent identity): it's pre-consolidation, structurally necessary before enterprises deploy agents at all, and identity layers historically produce durable standalone winners (Okta, CyberArk) rather than acquired features.
---
4. White Space
- Inter-agent trust and transaction brokering ("firewall between agents"): As enterprises deploy agents from multiple vendors that call each other, no one provides real-time policy enforcement, provenance, and authorization between autonomous agents mid-transaction — NHI players secure credentials at rest, not agent-to-agent negotiation at runtime.
- Cyber insurance underwriting for autonomous systems: No player offers continuous, attested measurement of "what your AI agents are permitted and empirically able to do" as an underwriting/liability standard. Whoever owns the actuarial data layer for agent-caused incidents (who's liable when an agent misfires or is hijacked?) sits at the intersection of a legal vacuum and a mandatory purchase.
Confidence note: company lists in Segments A, B, and E are high-confidence; acquisition details in Segment C are recent and worth verifying (Lakera/Check Point and Prompt Security/SentinelOne especially).