Open models. Criminal enablement. Autonomous systems.
XORTRON studies AI under adversarial conditions: modified safeguards, private deployment, operator control, tool access and autonomous execution. We build, test and publish systems to measure what changes when the assumptions behind a controlled deployment no longer hold.
Most AI risk discussions begin with the model as its developer intended it to behave. XORTRON starts where that assumption ends: open weights, altered safeguards, private inference, tool access and systems operated for adversarial objectives.
Model development, comparative evaluation and threat research are used together so capability is measured in context instead of inferred from policy language, refusal rates or product claims.
Follow open-weight families, modified releases, inference platforms, agent frameworks, public tooling and emerging misuse patterns.
Score the substance of the response. A warning followed by useful instructions is still useful; a refusal matters only when it actually withholds capability.
Create research models and deployment variants that make changes in alignment, capability, autonomy and behavior directly observable.
Release benchmarks, models, public deployments and research so the claims can be inspected, challenged and compared by others.
Engagements begin with a concrete intelligence or security question, then define the evidence, access, scope and output needed to answer it. No generic feed and no one-size-fits-all assessment.
Investigate adversary behavior, criminal ecosystems, tooling and AI-enabled misuse around a defined intelligence question, with findings tied to the actors, methods and exposure that actually matter.
Map exposed assets and relevant vulnerabilities, then separate theoretical risk from exposure an attacker could realistically use against the environment under review.
Investigate impersonation, phishing, identity abuse and deceptive infrastructure targeting a public presence, then organize the evidence for action, escalation or continued monitoring.
Trace vendor dependencies, exposed services and reported incidents to identify where third-party technology or suppliers create meaningful security questions for the organization.
Red-team models, agents and AI workflows against defined scenarios, including prompt injection, safeguard failure, information exposure, tool access and misuse potential.
Use AI as an analyst multiplier for research, evidence organization and investigative planning while keeping source access, verification and human review explicit.
Attributable HuggingFace model downloads across darkc0de & XORTRON related work.
Year the first XORTRON model was released publicly.
Open-weight research timelinePurpose-built benchmark families spanning enablement, autonomy and high-risk misuse.
Research evaluation suiteBrowser access to the XORTRON chat deployment and its research models.
Hugging Face SpacesXORTRON treats model behavior as an empirical problem. Capability, operator control, deployment conditions, tool access and autonomy are examined together because none of them tells the full story alone.
Measures how much a response could practically enable unlawful activity.
Studies delegation, sequencing, persistence and machine-speed execution risk.
Compares stock alignment, fine-tuning, abliteration and deployment posture.
Tracks emerging misuse patterns, public artifacts and criminal-AI narratives.
Judge the full operating system: the model, the operator, the deployment, the available tools, the permissions and the environment in which it acts.
Look past refusal language and safety framing to the planning, detail, adaptability and reliability contained in the answer itself.
Hosting, infrastructure control, tool access, memory and permissions can radically change what the same underlying model can accomplish.
Test how tool use, feedback loops, retries and persistent execution change speed, scale and the amount of human effort an operation requires.
Tie claims to observable behavior, comparative tests and public artifacts whenever possible so other people can inspect and challenge the result.
Purpose-built evaluations measure criminal enablement, autonomous behavior and general capability without reducing the result to whether the model printed the “right” safety language.
Structured evaluation of whether model output materially reduces the effort, expertise or uncertainty required to carry out criminal activity.
Examines the risk created when a model can plan, adapt, call tools, persist across steps and act with reduced human intervention.
Specialized benchmark tracks for severe abuse categories, designed to evaluate enablement while preserving research context.
Separates generic harmfulness from real operational assistance by testing planning depth, adaptability and usable detail.
Pairs safety testing with capability measurement so risk findings are interpreted against the underlying intelligence of the model.
Compares refusal behavior, reasoning quality, tool use and operational visibility across stock, modified, local and hosted deployments.
R.I.C.O. is XORTRON’s experimental adversarial model program: a working testbed for open-weight capability, alignment changes and criminal-enablement research. It exists so those questions can be tested on real systems instead of argued about in the abstract.
AI-enabled crime does not replace existing criminal methods; it plugs into them. The work follows those connections across identity abuse, ransomware, exposed infrastructure, criminal markets and autonomous systems.
Examine extortion activity, intrusion patterns and criminal incentives to support preparedness and investigative priorities.
Assess fraudulent messages, impersonation and deceptive web activity affecting organizations, customers and public identities.
Connect exposed assets and security findings with adversarial use cases and remediation priorities.
Analyze credential-abuse scenarios, synthetic identities and account-takeover risks within a defined investigative scope.
Evaluate prompt injection, information exposure, safeguard behavior and abuse pathways in models and AI applications.
Place vulnerability reports in context: affected technologies, potential exploitation and relevance to the environment under review.
Organize evidence, examine criminal methods and develop research questions that support defensible investigative conclusions.
Research accessible and authorized sources to evaluate criminal-market narratives, model offerings and emerging misuse patterns.
Assess dependency concerns and third-party incidents that may affect an organization’s wider security posture.
Connect adversary reporting, behavior patterns and investigative hypotheses to support a focused hunt.
Research and assessments for security, risk and fraud teams evaluating external threats, AI adoption and model-related exposure.
Evidence-driven research on criminal AI use, adversarial models, cyber-enabled crime and emerging digital threats.
Independent threat research, model evaluation and AI-risk analysis that can plug into existing security and investigative workflows.
Focused analysis of fraud, impersonation, identity abuse, synthetic media and technology dependencies affecting financial risk.
XORTRON Criminal Computing is an independent adversarial-AI and cyber-threat research operation based in Tampa, Florida. Since 2024, the project has grown across open-weight model development, public deployments, criminal-enablement benchmarks, autonomous-agent evaluation and published research, with more than 750,000 attributable model downloads across related Hugging Face work.
Run the deployment, examine the models, read the evaluations and compare the published research. XORTRON’s public artifacts are part of the evidence, not decoration around it.
Interact with the public XORTRON deployment and observe how the project behaves outside a static benchmark.
The Criminal AI Stack, Machine-Speed Intrusions and The AI Visibility Gap — merged into one 18-page XORTRON research collection.
Benchmark programs designed to measure practical enablement and autonomous-agent risk.
Three original XORTRON field reports in one continuous publication: The Criminal AI Stack, Machine-Speed Intrusions, and The AI Visibility Gap. Read the merged collection directly on the site without leaving the page.
Long-form analysis of AI-enabled crime, model evaluation, open systems and the gap between what a benchmark measures and what a deployed system can actually do.
Explores a possible convergence between AI-driven labor displacement and AI-enabled cybercrime: fewer entry-level security pathways, more technically capable people under economic pressure, and automation that can multiply the output of a single operator.
Read the article →
A lightweight evaluation method where a frontier model creates a fresh benchmark and private rubric, candidate models answer the same test, and the original frontier model scores the results. The method separates willingness, capability and real enablement.
Read the article →
Argues that agent performance should be judged by time-to-correct-result rather than raw intelligence or raw generation speed alone. Faster models can sometimes outperform stronger models when they can rapidly use objective feedback and retry.
Read the article →The project includes a growing technical reference library alongside the models and benchmarks, covering AI, machine learning, OSINT, security research, privacy, training and practical tooling.
A curated slice of the AI, security, OSINT, privacy and learning resources used around the project. The complete bookmark collection is included below.
XORTRON is an independent adversarial-AI and criminal-computing research operation based in Tampa, Florida. Its public work spans model development, criminal-enablement evaluation, autonomous-agent research, published analysis and a live AI deployment.
XORTRON offers research and assessment work across cyber threat intelligence, exposure management, digital risk, third-party risk and AI security. Each engagement is scoped around a specific question, the evidence available and the output needed to answer it.
XORTRON builds, deploys and evaluates the kinds of systems it studies. That makes it possible to compare capability, alignment, deployment and criminal enablement directly instead of relying only on vendor claims or refusal behavior. Related open-model work has passed 750,000 attributable downloads.
The scope includes ransomware, phishing, impersonation, credential abuse, exposed infrastructure, supply-chain risk and criminal markets, together with AI-specific issues such as prompt injection, model modification, tool access and autonomous misuse.
Security and fraud teams, researchers, government and law-enforcement analysts, managed security providers, financial organizations and others who need to understand adversarial AI beyond a product demo or policy sheet.
Use the live app embedded on this page, open the public deployment directly, browse the model collection, inspect the benchmarks and read the published research. The project is designed to be explored rather than taken on faith.
Run the system, inspect the model work, read the evaluations and use the research to judge the capability for yourself.
The full bookmark collection, right here: AI, OSINT, cybersecurity, privacy, learning and technical references.
1,602 links across 70 categories
No bookmarks match your search.
