The next advantage in cybersecurity belongs to teams that can turn machine intelligence into useful, repeatable security work. For a regulated financial business, that means more than finding a weakness. It means understanding what was tested, where sensitive information went, what the result proves and what to do next.
Decima is being built to put that entire process within reach of the regulated mid-market: security-specialized AI, control over deployment, an instrumented execution environment and reports people can act on.
Our position starts with a practical conviction: a security team should be able to choose its models, set its boundaries and retain the evidence. The architecture should support those choices as the technology changes.
The capability is real. Access is only part of the problem.
AI-assisted vulnerability research has advanced beyond summarizing scanner output. In its published evaluations, Anthropic reports that its Mythos Preview model found previously unknown vulnerabilities and developed exploits in established software. Those results illustrate the potential of agents equipped for sustained security work. Anthropic’s technical research.
Access is evolving too. Programs such as Anthropic’s Project Glasswing and OpenAI’s Daybreak provide advanced capabilities through controlled access, with eligibility and deployment conditions that vary by program. OpenAI explicitly describes reducing refusals for approved security work. Project Glasswing, OpenAI’s Daybreak update.
But access to a capable model does not give a credit union or growing financial platform a complete security operation. Someone still has to connect it to the right tools, define scope, supervise execution, validate results, manage cost and deliver evidence that engineers and executives can use.
That is the opportunity Decima is pursuing.
Model independence and control over sensitive data
A hosted model API can be an appropriate choice. Its suitability depends on the engagement’s data, contracts, retention settings, processing locations and access controls. Zero trust assesses access to resources without granting implicit trust based on network location; cloud use alone does not determine whether an architecture meets those principles. NIST’s zero-trust architecture guidance.
Some customers nevertheless need more control over sensitive source code, credentials, network details and assessment evidence. Decima’s intended architecture uses open-weight models served on infrastructure selected for the engagement. Dedicated rented compute is a practical starting point; deployment within a customer-controlled environment is a longer-term option to validate against that customer’s requirements.
Self-hosting can remove the need to send inference requests to a third-party model API. It does not automatically remove the infrastructure provider, secure every log or keep all processing within a national border. Those boundaries must be designed, documented and verified.
We also want to keep model choice open. New releases should be evaluated against real security tasks, with customers offered suitable options as our validated catalog develops. A larger model earns its place through better results at an acceptable cost. Parameter count alone is a poor buying criterion.
Security-specialized reasoning. Explicit execution boundaries.
General-purpose models can decline legitimate offensive-security tasks. That makes completion reliability a meaningful evaluation criterion alongside reasoning quality, tool use and accuracy. Our proposed model evaluations include security-tuned and refusal-reduced variants, sometimes described as abliterated models.
Removing refusal behavior is not evidence of security expertise. Each candidate still needs to demonstrate useful performance, including whether it follows task instructions, recognizes uncertainty, handles tool failures and produces claims that survive verification.
The operating boundary belongs in the platform. Written authorization, approved targets, exclusions, time and spending limits, tool permissions and stop controls must apply to every model. A model’s suggestion never grants itself permission to act.
This is what independence means for Decima: freedom to choose and improve the reasoning engine, within an execution environment whose authority is explicit and whose actions are observable.
A real working environment for the agent
Our proposed starting point is a custom, headless Kali Linux environment with a versioned security-tool inventory. Kali provides container bases and tool packages that can be assembled for this purpose. The model-serving system remains separate from the tool workers. Kali’s container documentation.
The division of work is deliberate. Models interpret observations, prioritize investigations and propose the next step. Versioned tools perform concrete checks. Decima’s harness coordinates that work, applies scope and budget controls, and records the commands, results and supporting artifacts.
For authorized source review, an agent could inspect repositories, run bounded local tests and investigate failed hypotheses. For an approved network assessment, workers would operate only within its permitted targets and techniques. These are distinct execution policies connected to the same evidence system.
Our first milestone is an actual instrumented run: enough retained detail to show agents working live, replay their recorded activity and analyze the results afterward. The dashboard should reveal the work, including failed attempts and observed defenses.
Evidence that survives a second look
A persuasive explanation is not proof of a vulnerability. A useful finding connects an affected system, the conditions needed, an observed result, its business impact and a practical fix. A missing protective layer is not automatically an exploitable weakness if another control prevents the attack.
Cloudflare’s published security-audit skill offers a useful reference for source review: separate discovery and verification roles, explicit coverage records and structured findings with distinct validation states. We are evaluating those patterns for Decima’s evidence pipeline. Cloudflare’s security-audit workflow.
Our evidence model must also account for the people operating the platform. Customer approvals, staff changes, model actions, tool execution, evidence access and report publication all need attributable history. Replaying a run should reconstruct what happened from saved records; repeating a test is a new operation.
Live exposure. Beautiful reports.
The report is a first-class product. We want the same reviewed evidence to support a live Operations view, an Assurance view for compliance teams and a polished PDF that stands on its own in a board pack or audit discussion.
That report should meet or exceed the presentation and analytical quality customers expect from leading cybersecurity firms: clear executive narrative, meaningful charts, reproducible findings, prioritized remediation, scope, limitations and relevant control mappings.
Daily report editions are part of the vision. Publication time and testing time must remain distinct: a new report can summarize existing evidence and remediation progress without implying that every system was retested that day.
Direct delivery to platforms such as Vanta and Drata is planned. Their documented APIs support evidence workflows, but a Decima connector still requires implementation, customer authorization and validation. An uploaded report does not automatically establish compliance or auditor acceptance. Vanta document API, Drata evidence API.
Continuous by design. Paced by your budget.
The ambition is ongoing validation as systems change. The customer’s budget, approved scope and operating requirements determine how frequently work runs.
A business might prioritize a payment API after releases, revisit critical exposures frequently and assess a wider estate on a slower schedule. The platform should make those tradeoffs visible: what is current, what is aging, what is queued and what another assessment would cost.
Repeated testing and expert human work can reinforce each other. The objective is to shorten the time between a material change, an actionable finding and a verified fix.
Built from the regulated operator’s perspective
Decima began with a need I encountered as CTO and co-founder of Byte Federal. Experience with financial infrastructure, GLBA obligations and ISO 27001 work shapes the questions we ask of an assessment: can we explain the exposure, reproduce the result, demonstrate remediation and account for who did what?
Our relationships across industry, cyber-forensics and universities inform the research direction. They are a reason to keep testing assumptions and evaluating new models as they emerge. Byte Federal’s compliance experience informs Decima; it is not a certification of the Decima product.
The advantage we are building is the combination: independent model choice, controlled execution, continuous evidence and excellent reporting, with service cadence shaped by the customer’s means.
We attack you first. Within your authorization, with evidence you can inspect and a path to making the next attempt fail.
Discuss a Decima pilot or explore our proposed workflow and reporting.