Dreadnode

• An agent-framework or coding-agent vendor could add security-specific capabilities and remove the reason to buy a separate security-agent platform.
• A security platform already in the buyer's stack could bundle agent building and evaluation, absorbing the use case Dreadnode sells on its own.
• The move from AI red teaming to general security-agent infrastructure could put the company against developer tooling on price, where its offensive-security reputation counts for less.
• With no named customer in the public record, paid adoption could prove thinner than the Series A syndicate implies.

Dreadnode sells the infrastructure for security agents rather than a finished agent. Its homepage argues that security is going through the transformation software engineering already went through, and that the machine-learning operations layer underneath it is missing. The about page states the same position, that the company builds the platform layer so teams can build and operationalize security agents on their own terms.

The buyer it addresses is the security team that wants to apply AI to offensive work. Documentation covers web-application testing, network operations and vulnerability research, and the quickstart walks a user from installing the command-line tool to running an authorized web penetration test that ends in a report.

Independent reporting frames the same problem from the research side. According to a Dark Reading article on the company's open-source release, its AI research scientist says models from frontier labs are better at offense than at defense and that generating high-quality training data for defensive tasks is difficult. That account comes from a Dreadnode researcher, so it records the company's own framing rather than an outside measure of buyer demand. [s1, s4, s6, s19, s10]

The platform covers the whole life cycle of a security agent. Public documentation describes a command-line tool and terminal interface, hosted evaluations, managed sandboxes, model training and prompt optimization, synthetic network-environment generation, a capability registry, tracing of every tool and model call, and self-hosting on the customer's own Kubernetes cluster.

AI red teaming is the deepest documented capability. The reference covers attack strategies, transforms and scorers spanning generative jailbreaks, adversarial machine learning, multimodal probing and multi-agent attacks, and it maps findings to OWASP, MITRE ATLAS, NIST AI RMF and Google SAIF for export as reports.

Release notes record a weekly cadence. Seven consecutive weekly entries, dated July 2 to August 13 on the feed, cover a Claude Code engine, end-to-end multimodal red teaming, an agent memory and session-monitoring overhaul, a traditional-machine-learning red-teaming suite, and a versioned registry for structured agent output.

Some of the engineering is open to inspection. Dreadnode publishes a lightweight framework for interacting with language models, and example agents, on GitHub, and released DreadGOAD, a reproducible Active Directory training environment, and Ares, a system that runs red and blue agents against the same live environment. [s2, s6, s9, s18, s3, s1]

Dreadnode names the two rivals it positions itself between. Its 2.0 announcement contrasts its approach with closed vendors that deliver a single-purpose agent and with generic agent frameworks that were not built for security's requirements, and claims the middle as security-specific infrastructure.

That middle is a wider arena than the one the company started in. A pitch to be infrastructure for any security agent puts the product beside the agent tooling developers already use and beside security platforms that could add agent building to what they sell.

The claim Dreadnode leads with is breadth. Its own pages describe build, evaluate, train and deploy in one platform, and the reviewed sources establish that the platform covers those stages without establishing that a buyer chose it for that reason. [s3, s2]

Dreadnode runs a self-serve motion with an enterprise tier above it. The Pro plan starts free with $25 of complimentary credits, bills pay-as-you-go against credits worth a cent each, and charges nothing for seats. The Enterprise plan adds a one-time annual fee, on-premises deployment, dedicated engineer support and custom capability work.

No source reviewed for this catalog names a customer. The site's sitemap lists 42 pages and includes no customers page, and the company's own Series A announcement describes government, enterprise and AI-lab partners it does not name. The visible backing is instead the $14 million Series A from Decibel, Next Frontier Capital, In-Q-Tel, Sands Capital and Indie VC, with Aviso Ventures named among the investors as well.

What the company puts in front of practitioners is research and events. It jointly organized the 2025 Offensive AI Con with RemoteThreat and DevSec, and its July 2026 announcement lists an accepted Black Hat USA briefing, a workshop at a Las Vegas AI security forum and talks across DEF CON villages. That program puts the company in front of practitioners who could buy the self-serve plan themselves, and no reviewed source records one who did. [s5, s17, s21, s11, s4, s7, s20]

Dreadnode's about page lists Brad Palm as chief executive, with Will Pearce as co-founder and Nick Landers as co-founder and chief technology officer. The March 2026 launch release quoted Pearce as CEO and co-founder, so the chief-executive title moved after that announcement.

Press corroborates the founders' pedigree. SecurityWeek describes Pearce as the former lead of NVIDIA's AI red team and Landers as a former VP of Research at NetSPI.

The wider bench publishes under its own name. Company researchers wrote the AIRTBench benchmark paper and the Crucible attack-data analysis, introduced the DreadIndex model index, had an accepted Black Hat USA briefing on the August 6 program, which a Dark Reading article describes, and announced an AI Village presentation of an agent-failure-modes paper co-authored with a Google DeepMind researcher. Dreadnode states the team draws on operators from NVIDIA, Microsoft, NetSPI, Meta, Cohere and Run Sybil. [s4, s3, s11, s12, s7, s10, s14, s15, s8]

Dreadnode's clearest readiness signal is deployment choice. The Enterprise plan offers on-premises deployment, the documentation covers self-hosting on the customer's own Kubernetes cluster, and the pricing page states that customer data is never used for training and that evaluation, training, Worlds and AI red teaming data stays inside the customer's environment.

The company publishes a trust center on a Vanta-hosted subdomain. It lists controls across infrastructure, organizational, product, internal-procedure and data-privacy categories, names Amazon Web Services, Cloudflare, Google Workspace and Vanta as subprocessors, lists one resource, an engagement letter, and carries a request-access control.

No completed attestation appears in the reviewed sources. The rendered trust center lists no SOC 2 or ISO 27001 report, the probed homepage, about and pricing pages carry no attestation wording, and the published sitemap lists no compliance page. For an enterprise motion that includes on-premises deployment, a buyer would still ask for an audit report before purchase. [s5, s2, s16, s17]

Add analyzed competitors to compare them side by side with Dreadnode.

Dreadnode addresses two buyer tiers from one price list. The Pro plan is aimed at builders, practitioners and teams and starts free with complimentary credits, while the Enterprise plan is aimed at organizations that need scaled agent workloads, on-premises deployment and custom capabilities.

The personas follow the agent life cycle. The user is the offensive engineer or red teamer running capabilities against authorized targets, the champion is the security engineer or model developer standing AI up for security work, and the enterprise buyer owns the deployment and data-control requirements the higher tier answers.

Two signals suggest government and regulated buyers without confirming one. In-Q-Tel joined the Series A, the enterprise tier sells on-premises deployment and customer-held data, and the company's research program runs on Active Directory environments typical of large organizations. No reviewed source names a public-sector or regulated customer, so segment priorities read from the product and the price list rather than a disclosed roster.

The platform covers a security agent's whole life cycle. Documentation and the 2.0 announcement describe a command-line tool and terminal interface, hosted evaluations, managed sandboxes, model fine-tuning and prompt optimization, synthetic network-environment generation, a resource hub of roughly 1,600 security tasks with packaged capabilities, and tracing that records each tool and model call.

The deepest documented capability is AI red teaming. The reference covers attack strategies, transforms and scorers spanning generative jailbreaks, adversarial machine learning, multimodal probing and multi-agent attacks, with findings mapped to OWASP, MITRE ATLAS, NIST AI RMF and Google SAIF for export as reports.

The advantage the company presses is offensive expertise turned into measurement. Its published benchmark work measures how language models autonomously discover and exploit weaknesses in AI systems, drawing on 70 capture-the-flag challenges from the Crucible environment and an analysis of 214,271 attack attempts by 1,674 users, and its DreadIndex index scores models across 76 offensive-security tasks.

The design choice is horizontal infrastructure rather than a finished agent. The customer brings the domain expertise while the platform supplies the build, evaluate and deploy machinery, which the quickstart makes concrete: install the command-line tool, add the web-security capability, point it at an authorized target and let the agent produce a report.

Dreadnode runs a product-led motion with an enterprise tier above it. The Pro plan starts free with $25 of credits and bills pay-as-you-go with no commitment, so a practitioner can adopt before any sales contact, and the Enterprise plan moves to a quoted annual arrangement with on-premises deployment and dedicated engineer support.

No reviewed source names a customer. The site's sitemap lists 42 pages and includes no customers page, and the company's own Series A announcement describes government, enterprise and AI-lab partners it does not name. The visible backing is instead the $14 million Series A from Decibel, Next Frontier Capital, In-Q-Tel, Sands Capital and Indie VC, with Aviso Ventures named among the investors as well.

Dreadnode generates demand through research and events. The company jointly organized the 2025 Offensive AI Con with RemoteThreat and DevSec, its July 2026 announcement lists an accepted Black Hat USA briefing, a workshop at a Las Vegas AI security forum and talks across DEF CON villages, and it publishes benchmarks and open-source tools that put its work in practitioners' hands directly. Those practitioners are the people the free self-serve plan is priced for, and no reviewed source records one of them buying it.

Dreadnode prices by consumption rather than by seat. The Pro plan bills credits worth a cent each against inference and compute, charges no monthly fee, and puts no separate price on seats, so a team can add users without adding cost.

The enterprise arrangement adds a fixed component to the same meter. A one-time annual fee varies with deployment requirements, service levels and custom work, and usage-based credits continue on top. Dreadnode publishes both structures openly, which fits a self-serve-first stance.

Dreadnode ships the platform both as managed software and as software the customer runs. The Pro plan is fully managed, the Enterprise plan supports on-premises deployment with the customer controlling its own data, and the documentation covers installing the platform on the customer's own Kubernetes cluster.

Dreadnode concentrates operational control in the managed sandbox. Segmented, scalable compute runs agents with guardrails built in, hosted evaluations absorb the rate limits, timeouts and monitoring that agent testing otherwise imposes, and tracing records every tool and model call so a team can reconstruct what an agent did.

The release feed shows an operating cadence. Seven consecutive weekly entries, dated July 2 to August 13 on the feed, cover a Claude Code engine, end-to-end multimodal red teaming, an agent memory and session-monitoring overhaul, a traditional-machine-learning red-teaming suite, and a versioned registry for structured agent output.

Dreadnode's readiness shows most in how it handles the customer's data. The Enterprise plan offers on-premises deployment, the documentation covers self-hosting on the customer's own Kubernetes cluster, and the pricing page states that customer data is never used for training and that evaluation, training, Worlds and AI red teaming data stays inside the customer's environment.

The company publishes a trust center on a Vanta-hosted subdomain. It lists controls across infrastructure, organizational, product, internal-procedure and data-privacy categories, names Amazon Web Services, Cloudflare, Google Workspace and Vanta as subprocessors, lists one resource, an engagement letter, and carries a request-access control.

No completed attestation appears in the reviewed sources. The rendered trust center lists no SOC 2 or ISO 27001 report, the probed homepage, about and pricing pages carry no attestation wording, and the published sitemap lists no compliance page. For an enterprise motion that includes on-premises deployment, a buyer would still ask for an audit report before purchase.

Dreadnode frames the product as infrastructure other security work runs on. The 2.0 release positions it as a platform for security agents rather than a point tool, with a resource hub where teams publish and version capabilities privately or share them with a community.

Part of the ecosystem is deliberately open. The company publishes a lightweight framework for interacting with language models, and example agents, on GitHub and released DreadGOAD and Ares, an Active Directory environment and a red-versus-blue agent system that other teams can run against their own agents.

The platform frame also widens the field of rivals. Sitting between closed single-purpose agents and general agent tooling means the tooling a developer already uses competes for the same build-and-deploy work, and no reviewed source shows a community large enough to hold the platform in place.

Dreadnode's about page lists Brad Palm as chief executive, with Will Pearce as co-founder and Nick Landers as co-founder and chief technology officer. The March 2026 launch release quoted Pearce as CEO and co-founder, so the chief-executive title moved after that announcement.

The founders' backgrounds are corroborated by press. SecurityWeek describes Pearce as the former lead of NVIDIA's AI red team and Landers as a former VP of Research at NetSPI.

The wider bench publishes under its own name. Company researchers wrote the AIRTBench benchmark paper and the Crucible attack-data analysis, introduced the DreadIndex model index, had an accepted Black Hat USA briefing on the August 6 program, which a Dark Reading article describes, and announced an AI Village presentation of an agent-failure-modes paper co-authored with a Google DeepMind researcher. Dreadnode states the team draws on operators from NVIDIA, Microsoft, NetSPI, Meta, Cohere and Run Sybil.