Expertise that grows.
Each investigation builds on codebase knowledge, security patterns and lessons from previous research.
A STARTUP IN STEALTHRESEARCH PREVIEW
CYBERGYM SUCCESS RATE 98.74%Latest evaluation Publication in progress
VERIFIED VULNERABILITIES 26Discovered across critical open-source infrastructure
PRIVATE. SOVEREIGN. MODEL-NEUTRAL.
Cyber AI.
Under your control.
Advanced vulnerability research.
Your models. Your infrastructure. Your expertise.
FOR TECHNOLOGY PARTNERS, INVESTORS & SOVEREIGN AI PIONEERS
01 / ABOUT
A new foundation
for Cyber AI.
CybAgents is a stealth startup developing private, sovereign and model-neutral Cyber AI for governments and enterprises.
Our proprietary cognitive layer adds persistent cyber knowledge and research expertise to the models you approve. Our first application: software vulnerability research.
Freedom to choose.
Choose qualified frontier, European or open-weight models. Deploy in your private cloud, on-premises or an air-gapped environment.
Intelligence you retain.
Your cyber expertise evolves independently of model releases. Your code and research context stay under your control.
CYBAGENTS ARCHITECTURE
Frontier modelsOpen-weight modelsPrivate modelsSovereign AI
CybAgents Intelligence layerCyber knowledgeCognitive reasoningCodebase memory
Your enterprise codebaseWITHIN YOUR TRUST BOUNDARY
01 / 03
Start with a model approved by your organization.
02 / CYBERGYM
Research capability.
Measured in execution.
Developed by UC Berkeley researchers, CyberGym tests AI agents on real software vulnerabilities. Success requires a working reproduction.
LATEST CYBAGENTS EVALUATIONSEPT 2026
98.74%
Success rate · pass@1
Updated first-place performance
Company-reported result. Official submission and publication in progress.
1,507Real vulnerability tasks
188Software projects
CyberGym Level 1 supplies an unpatched codebase and a vulnerability description. Agents generate a test that reproduces the known flaw. New vulnerability discovery is a separate research result, presented below.
03 / VULNERABILITY RESEARCH
Beyond the benchmark.
Real discoveries.
26 individually verified vulnerabilities across four critical open-source projects.
Our research portfolio includes memory-safety flaws, logic errors and cross-subsystem issues, supported by sanitizer output, reproducible tests or execution evidence.
OPERATING SYSTEMS14findings
Linux kernel
Memory-safety and protocol-handling flaws across networking, filesystems and asynchronous I/O.
SELECTED OUTCOME
A FUSE finding in the Linux subsystem that enables filesystems to run in user space.
NETWORK SERVICES5findings
dnsmasq
Memory-safety and input-handling flaws in widely deployed DNS and DHCP components.
DISCLOSURE
Findings reported to the maintainer. Remediation status varies by issue.
FILE SERVICES3findings
Netatalk
Memory corruption and integer-handling flaws in network file-sharing software.
RESEARCH EVIDENCE
Reproducible impact demonstrated through controlled execution.
TRANSFER SERVICES4findings
ProFTPD
Arithmetic, logic and stability flaws in file-transfer and extension-handling code.
RESEARCH EVIDENCE
Four findings confirmed through the research pipeline.
ADDITIONAL APPSEC RESEARCH
Discuss the research The same approach, applied to enterprise applications.
Research campaigns also covered Keycloak, Grafana, Gitea and other widely used applications, with 14 targets audited and 12 verified exploits documented.
CybAgents research portfolio. The 26 findings cover C/C++ infrastructure research, including a parallel discovery; AppSec results are separate. These summaries do not imply that every finding is publicly disclosed or fully remediated.
04 / SOLUTIONS
CodeVR.
Private AI vulnerability research.
For your most sensitive code.
FIRST PRODUCT / PREVIEW
CodeVR brings our research methodology to AppSec, product security and engineering teams.
Discover, validate and help remediate complex vulnerabilities, with codebase-specific memory and customer control.
Explore an early evaluationDEEP VULNERABILITY DISCOVERY
Follow the logic.
Find the hidden risk.
Investigate relationships across your codebase, reason about complex bug classes and build on knowledge of your architecture.
Codebase context / Vulnerability patterns / Agent researchRESEARCH PATHIllustrative workflow
Application entry points01Cross-module relationships02Vulnerability hypothesis03EXPLOITABILITY VALIDATION
Turn a hypothesis
into evidence.
Investigate whether a suspected vulnerability can be reproduced. Give security teams evidence to assess the finding and prioritize their response.
Reproduction / Execution feedback / Exploitability reasoningVALIDATION PATHIllustrative workflow
Candidate vulnerability01Controlled reproduction02Evidence for review03SECURE REMEDIATION SUPPORT
Make the next
fix more informed.
Bring actionable remediation suggestions into the developer workflow. Preserve what the team learns in codebase-specific memory.
Remediation playbooks / Human review / Persistent contextREMEDIATION PATHIllustrative workflow
Validated finding01Remediation guidance02Review & codebase memory03SOVEREIGNTY ACROSS THE STACK
Keep code, models and memory
inside your trust boundary.
Qualify the model for capability, cost and jurisdiction. Keep human oversight and traceable research actions within the environment you approve.
Inside your perimeter.
Designed for customer-controlled infrastructure, including on-premises and air-gapped deployment requirements.
Private, on your terms.
Bring the cyber intelligence layer into your approved private cloud, with the models and controls your organization selects.
Aligned with your jurisdiction.
Designed for sovereign cloud environments where data location, model access and infrastructure independence are strategic priorities.
YOUR INFRASTRUCTURE. YOUR CONTROL.A BROADER RESEARCH HORIZON
Code first.
More security domains to follow.
The cognitive layer can extend into other security workflows by adding domain knowledge, tools and dedicated evaluation.
Future directions. Further applications and their results will be announced as the research progresses.
05 / CONTACT US
Still in stealth.
Open to the right
conversations.
CybAgents is engaging with technology partners, investors and selected enterprise and government teams interested in exploring sovereign AI for vulnerability research.
Technology partners
Evaluate our cyber layer with your model or infrastructure.
Investors
Discuss the technology, research evidence and opportunity.
Enterprise & government
Explore a scoped evaluation on your code and trusted stack.
Start a conversation
Tell us a little about your interest in CybAgents.