The future of cybersecurity is AI vs. AI.
XORTRON Criminal Computing conducts research into Artificial Intelligence (AI) Offensive Security and Vulnerability Management. We investigate AI misuse, agentic abuse and harmful applications of AI across models, agents, APIs and exposed systems through adversarial evaluation, vulnerability analysis and cyber-threat research. Findings are tied to evidence, security impact and practical remediation priorities. Live, in-person and virtual demonstrations of criminal AI in action are available on request.
Security risk is defined by what can be exploited, under which conditions and with what impact. XORTRON studies model behavior alongside deployment architecture, tool permissions, public-facing exposure, API abuse and automated exploitation.
Our OSVM approach connects offensive testing, vulnerability intelligence and adversarial AI evaluation. Analysis separates demonstrated weaknesses from theoretical concerns and identifies the conditions that would make an attack feasible.
Identify exposed services, relevant vulnerabilities, AI integrations, agent permissions and documented attacker methods.
Examine whether a weakness creates a plausible attack path, including required access, technical constraints and likely consequences.
Compare base and modified models, deployment configurations, prompt-injection exposure, tool misuse and autonomous-agent behavior.
Record evidence, scope, uncertainty and potential impact. Publish selected methods and findings for independent review.
Research covers AI attack surfaces, exposed technologies, vulnerability validation and documented threat activity. These topics inform our published work, model evaluations and live demonstrations.
Track adversary techniques, criminal infrastructure and AI-enabled attack methods. Map reported activity to affected technologies and credible exposure.
Analyze disclosed vulnerabilities, exposed services, exploitation prerequisites and attack paths to inform risk-based prioritization.
Examine phishing infrastructure, impersonation, credential theft and related abuse. Connect indicators to documented campaigns and affected surfaces.
Review software dependencies, public vulnerability records, third-party incidents and internet-facing supplier exposure.
Test AI models and agents against defined adversarial scenarios: prompt injection, API abuse, data exposure, safeguard failure, tool misuse and unauthorized actions.
Use AI to accelerate technical triage, reporting and intelligence correlation; verify consequential findings against primary evidence.
Attackers are using AI to automate reconnaissance, scale deception and find weaknesses faster. Defenders need AI capable of anticipating those same tactics—not an assistant that stops thinking where the threat begins.
Email info@xortron.ai to request a live, in-person demonstration in the Tampa area or a virtual demonstration. Tell us which format you prefer and what capabilities you want to see.
Attributable Hugging Face downloads across related darkc0de and XORTRON model releases.
First public XORTRON model release.
Open-weight research timelineBenchmark families covering model capability, offensive assistance, autonomy and high-risk misuse.
Research evaluation suitePublic AI deployment with associated model releases and evaluations.
Hugging Face SpacesXORTRON's research and development began in 2023, with the project's first language models publicly released on Hugging Face in 2024. Since its inception, XORTRON has pursued the development, evaluation, and advancement of unrestricted artificial intelligence, establishing a notable history of competitive performance on the Uncensored General Intelligence (UGI) Leaderboard, one of the most prominent and widely followed community benchmarks dedicated to uncensored and unrestricted large language models.
In 2025, darkc0de/XortronCriminalComputingConfig, a 24-billion-parameter model, reached the top of the UGI leaderboard among models under 70 billion parameters in both UGI and W/10 performance. Remarkably, from its release on April 30, 2025, until October 23, 2025, no other model in that parameter class surpassed its recorded performance on both metrics simultaneously. This nearly six-month achievement demonstrated XORTRON's ability to produce highly competitive unrestricted language models without relying on the enormous parameter counts of much larger systems.
That trajectory continued into 2026 with the release of darkc0de/XORTRON.CriminalComputing.LARGE.2026.3, a 123-billion-parameter model evaluated on February 23, 2026. Achieving a score of 66.16 UGI, the model debuted at #2 overall on the leaderboard, outperforming nearly every evaluated model regardless of architecture, parameter count, or whether its weights were publicly available. More than seven months later, as of October 10, 2026, it continues to rank #4 overall, #1 among non-thinking models, and #1 among open-weight models under 350 billion parameters, maintaining its position among the leaderboard's highest-performing systems despite continued advances in the broader AI industry.
XORTRON's technical achievements have been accompanied by growing community engagement. As of October 10, 2026, UncensoredChat, XORTRON's publicly accessible AI assistant, ranked #29 among all Hugging Face Spaces for likes gained during the preceding 7 days and #31 for likes gained during the preceding 30 days, according to Model Pulse. These rankings place XORTRON's application among the platform's most actively liked Spaces across both weekly and monthly reporting periods, reflecting significant community interest in the project's publicly available AI systems.
R.I.C.O. is XORTRON’s experimental open-weight model program for adversarial AI security research. It examines how model modification, distillation, reduced safeguards, inference configuration and agent tooling affect offensive capability, misuse exposure and the reliability of security boundaries.
XORTRON develops and tests open-weight models, evaluates autonomous agents and investigates AI-enabled crime, agentic abuse and other harmful applications of AI. Analyses account for system permissions, tool integrations, repeatability and real exposure—not only a model’s willingness to respond.
Measures operational assistance, specificity and accuracy in defined criminal-enablement scenarios.
Tests agent planning, tool invocation, persistence and execution across adversarial tasks.
Compares base models, modified checkpoints and security controls under controlled prompts.
Tracks documented AI-enabled intrusion methods, exploitation trends and threat activity. Explore the ORCA incident archive below.
An effective assessment accounts for the model, its hosting environment, accessible data, tool permissions, attacker access and potential impact.
Measure whether responses are accurate, usable, adaptable and repeatable under defined test conditions.
Check which system permissions, credentials, tools and deployment choices make a weakness actionable.
Test how repeated attempts, feedback, tool access and autonomous execution change attack feasibility.
Record methods, observed results, limitations and relevant comparisons so findings can be independently assessed.
XORTRON uses targeted evaluations to measure model assistance, agent execution, safeguard weaknesses and high-risk misuse. Scores are interpreted alongside the test conditions and observed outputs.
Measures whether outputs materially reduce the work or expertise needed for specified criminal tasks.
Measures multi-step planning, tool use and autonomous completion under defined misuse scenarios.
Tests selected high-consequence misuse scenarios with explicit limitations on interpretation.
Separates generic responses from actionable assistance using specificity and practical utility measures.
Uses general capability results as a reference point for interpreting offensive-assistance scores.
Compares safeguards, inference settings and tool permissions to identify changes in security behavior.
Explore documented AI agent incidents through ORCA's searchable database, monthly trend charts, severity assessments, attack surfaces, and list or table views. The archive is maintained in its own application and displayed here live.
Research covers disclosed vulnerabilities, exposed infrastructure, credential abuse, phishing, ransomware, automated exploitation and AI-enabled intrusion techniques. Findings focus on prerequisites and practical security consequences.
Track ransomware techniques, access vectors and extortion infrastructure with relevance to exposed assets.
Investigate phishing delivery, impersonation patterns and infrastructure associated with credential theft.
Link internet-facing exposure to relevant weaknesses, exploitation conditions and remediation priorities.
Study account takeover, stolen credentials and identity abuse in specific documented scenarios.
Test prompt injection, data leakage, model safeguards and tool-access boundaries in AI systems.
Correlate CVE disclosures, affected components, exploitation evidence and potential exposure.
Document AI-enabled fraud, criminal techniques, indicators and links between attacks, tooling and infrastructure.
Monitor lawfully accessible reporting and sources on criminal markets and illicit AI tooling.
Map dependency and vendor risk using published vulnerabilities and documented incidents.
Translate observed tactics and published intelligence into focused hunting hypotheses.
Technical analysis of exposure, AI security weaknesses and attacker methods for enterprise risk and security teams.
Research on AI-enabled crime, adversarial systems and documented attack methods for public-sector analysis.
Adversarial model assessments, vulnerability context and threat intelligence for managed security operations.
Investigation of fraud techniques, synthetic identities, credential abuse and AI-enabled impersonation.
XORTRON Criminal Computing is an independent AI Offensive Security and Vulnerability Management (OSVM) organization based in Tampa, Florida. Research began in 2023, followed by the first public model releases in 2024. Its work includes AI misuse investigations, offensive AI evaluation, vulnerability and exposure intelligence, criminal-enablement benchmarking and open-weight model development. Related model releases have recorded more than 750,000 attributable Hugging Face downloads.
Inspect released models, published evaluations, reports and the live UncensoredChat deployment. The results and artifacts are available for direct examination.
Use the public model deployment and compare observed behavior with published benchmarks.
The Criminal AI Stack, Machine-Speed Intrusions and The AI Visibility Gap — merged into one 18-page XORTRON research collection.
Benchmarks measuring model-generated offensive assistance and autonomous-agent behavior.
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.
Analysis of AI-enabled cybercrime, model capabilities, offensive evaluation methods and operational security.
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 →Reference tools for offensive security research, vulnerability intelligence, AI testing, OSINT and technical training.
Selected references for vulnerability research, security testing, model evaluation and open-source intelligence. The complete bookmark directory follows.
The future of cybersecurity is AI vs. AI. For offensive security, vulnerability management, AI red teaming and exposure analysis informed by today's cybercriminal landscape, contact info@xortron.ai. Tell us what needs protecting, what is in scope and what you need tested. All offensive assessments require authorization.
Directory of AI, OSINT, vulnerability research, cybersecurity and technical reference links.
1,602 links across 70 categories
No bookmarks match your search.
