Autonomous agent swarms breach 100+ sites, and cheap open models match gated ones on offence

The Continuous Agent Pentest idea has essentially been taken. XBOW, Horizon3 NodeZero, Hadrian, Pentera, Bugcrowd Pathseeker and YC startups like Parameter already sell continuous autonomous external testing ranked by exploitability, and Horizon3 in particular already sells to mid-market companies. Using open models is a cost advantage, not a selling point buyers care about. A new entrant could only win on that side by going narrow: owning manufacturing-specific external exposure such as internet-facing OT, remote access gateways, vendor VPNs and supplier portals. It would also need to bundle fixes or a managed service for teams with only two or three security staff, priced well below the incumbents. Even then, the go-to-market looks hard. The underwriting idea is less contested. Insurers today rely on passive outside-in scanning (Coalition, Corvus, SecurityScorecard-style ratings, Panorays) and on enrichment with exploit intelligence (Cytora with VulnCheck, Cyberwrite). I found no insurer or MGA partnership that uses agent-based exploitation to score a policyholder before binding or at renewal. The real opening is safe, authorized and fast proof of exploitability as an underwriting signal, plus loss data showing it predicts claims better than passive ratings. The obstacles are getting each applicant's legal consent to be attacked, the liability if a test breaks something during underwriting, and Onda's stated preference for quick insurability calls over deep assessment. A further risk is that an incumbent such as XBOW, Horizon3 or Coalition adds an insurer offering once the idea is proven.

• XBOW

  Autonomous AI hacker that tests applications continuously. Raised a $120M Series C at a valuation above $1B in March 2026, then added $35M from strategics in May. Aimed at enterprises, not the mid-market.

• Horizon3.ai NodeZero

  Mature autonomous pentesting platform covering internal, external, cloud and identity. Already sold into mid-market companies with small security teams, so it is the closest direct competitor to the Continuous Agent Pentest idea.

• Hadrian (Nova)

  Agentic pentesting focused on the external attack surface, with continuous asset discovery and exploit chaining. Overlaps almost completely with weekly external testing ranked by exploitability.

• Bugcrowd Savant Pathseeker

  Bugcrowd's agentic product for continuous testing of the external attack surface with proof of exploitability. In early access since July 2026, with general availability planned for later in 2026.

• Parameter (fka Hex Security, YC W26)

  YC startup whose AI agents run continuous pentests against apps and infrastructure, chain exploits and return reproduction steps.

• Pentera, FireCompass, Aikido and others

  Other vendors that 2026 roundups list as autonomous or continuous pentest options. This confirms the pentest category is crowded.

• Coalition (Active Risk Platform)

  Cyber insurer that scans policyholders' external footprint and reportedly can require fixes before renewal. Corvus, Beazley and AIG are said to do similar outside-in scanning. These are passive scans, not exploit attempts.

• Cytora + VulnCheck

  April 2026 partnership that adds vulnerability and exploit intelligence to cyber underwriting submissions. This is the closest existing underwriting input to an exploitability score.

• Onda + Panorays

  Cyber MGA that uses Panorays' external risk assessment to decide quickly whether a business is insurable. The MGA says it wants quick insurability calls rather than deep assessments of defences.

• Cyberwrite (with Markel)

  AI-driven cyber risk analytics used by Markel in Europe to price exposures. Another incumbent data source competing for insurer underwriting budgets.

Checked 2026-10-21