-
-
Vardryn logo: a shield-based identity for a governed AI cybersecurity response agent.
-
Audit-style incident report showing malware beaconing, severity, safety gate, and response tier.
-
WebShield scan showing critical website findings, quarantine, ticket creation, and no live file changes.
-
Red-team suite passing prompt injection, fake approval, secret leakage, and containment bypass tests.
-
Security Guard scan checking secrets, policy safety, permissions, and simulation-first protections.
Inspiration
I built Vardryn because I wanted to explore what a cybersecurity agent would look like if safety came first. A lot of AI projects focus on making agents more autonomous, but in cybersecurity, too much autonomy can be dangerous. An agent that acts too quickly could delete evidence, expose secrets, or take the wrong action based on weak signals.
I was also inspired by real-world website compromise scenarios, especially fake reCAPTCHA malware that tricks users into copying commands into Terminal, CMD, or PowerShell. That made me think about how small organizations, schools, and local businesses may not always have a full security team ready to investigate these threats quickly.
Vardryn became my answer to the question: how can we make AI useful for incident response without making it reckless?
What it does
Vardryn is a simulation-first AI cybersecurity response agent. It detects suspicious activity, creates audit-ready tickets, preserves evidence in sandbox quarantine, routes alerts through Slack, Discord, GitHub Issues, and email, and blocks unsafe actions using access control, an Agent Action Firewall, prompt-injection detection, and human approval.
It can also scan copied website folders through WebShield to look for fake reCAPTCHA patterns, suspicious JavaScript, clipboard hijacking, and other signs of website compromise without modifying live files.
How I built it
I built Vardryn as a local Python project with separate modules for detection, quarantine, alert routing, access control, red-team testing, prompt-injection defense, and agent tracing. I kept the system modular so each part could be tested without breaking the rest of the agent.
The project includes automated unit tests, scenario tests, and red-team simulations to check that Vardryn still behaves safely when it sees adversarial input, fake approvals, poisoned intel, or attempts to bypass human review.
Challenges
One of the biggest challenges was making the agent powerful without making it unsafe. I wanted Vardryn to detect and respond to suspicious activity, but I did not want it to take destructive actions, modify live systems, or trust weak evidence blindly.
Another challenge was designing the agent so it could explain what happened. That led me to add audit logs, local tickets, sandbox quarantine, an Agent Trace black-box recorder, and an Action Firewall that can decide whether an action should be allowed, denied, simulated, or sent for human review.
What I learned
I learned that building a cybersecurity agent is not just about detection. It is also about trust, evidence, permissions, and control. An AI agent needs to explain what it saw, what it decided, what it refused to do, and what still needs human review.
I also learned how important it is to test against unsafe inputs. Red-team scenarios helped me think beyond normal use cases and consider how an attacker might try to manipulate the agent through logs, website text, or fake approval messages.
What's next
The next step for Vardryn is to turn it from a strong local prototype into a more complete enterprise-ready AI security platform. My immediate plan is to add an Evidence Trust Score so Vardryn can explain not only that an incident is severe, but also how strong the supporting evidence is and what evidence is still missing.
After that, I plan to build an Evidence Integrity Layer with hash-chain audit logs so tickets, quarantine records, and incident reports can be checked for tampering. I also want to add a Network Evidence Analyzer for saved PCAP, DNS, firewall, or Zeek-style logs, plus a Static Malware Triage module that can safely inspect suspicious files without executing them.
Another major goal is to expand Vardryn’s intelligence layer. Approved threat intel should not just sit as passive context; it should be promoted through controlled levels such as enrichment, detection candidate, and active detection rule only after human approval, source trust checks, freshness checks, corroboration, and prompt-injection scanning.
I also plan to build a Streamlit dashboard so users can see incidents, evidence scores, quarantine sessions, red-team results, action firewall decisions, agent traces, and security readiness in one organized interface. Longer term, I want to add an Adversary Simulation Lab that safely simulates reconnaissance behavior, such as port scanning, DNS lookups, and web fingerprinting, using only controlled local fixtures and generated logs.
The overall goal is to keep making Vardryn more powerful while preserving its core principle: simulation-first, human-review-first, non-destructive, and governed by safety gates.
Log in or sign up for Devpost to join the conversation.