Skip to content
DeployWorks
한국어

Service

AI & Cyber Security

Threat modelling, prompt-injection defences, red-teaming and secure deployment for AI systems.

Start a conversationHow we engage

AI systems add new ways in, and your security team has to sign off.

LLM applications read untrusted text, call tools and touch sensitive data. That brings risks traditional reviews don't cover well: prompt injection, data leaking through answers, over-privileged agents and model supply chains. Security teams are asked to approve systems they have no checklist for.

We threat-model AI systems, test them the way an attacker would, and build the controls into the design, so the security review rests on evidence.

You may need this if

  • A security review is blocking an AI launch.
  • Your assistant or agent reads emails, documents or web pages from outside.
  • Agents can call internal tools or APIs.
  • You need to show AI controls to auditors or customers.

What we deliver

  • 01

    Threat model

    Data flows, trust boundaries and abuse cases for your AI system, each mapped to a concrete control.

  • 02

    Red-team testing

    Prompt injection, data exfiltration, jailbreak and tool-misuse tests, with reproducible cases.

  • 03

    Guardrails and least privilege

    Input and output filtering, scoped tool permissions and approval steps for risky actions.

  • 04

    Secure deployment

    Secrets handling, network policies, model and dependency provenance, and audit logging.

  • 05

    Security test suite

    Attack cases that run in CI, so fixes stay fixed.

  • 06

    Review pack

    Documentation your security team and auditors can review and keep.

Reference architecture

A reference design, not a fixed product. Discovery adapts it to your systems and constraints.

Controls around an AI application

  1. Input boundary

    • Authentication
    • Input filtering
    • Untrusted content marked

  2. Model and orchestration

    • Prompt policy
    • Model gateway
    • Rate limits

  3. Tools and data

    • Scoped tool permissions
    • Human approval
    • Permission-aware retrieval

  4. Output boundary

    • Output filtering
    • Sensitive data checks
    • Citations

Assurance

  • Threat model
  • Red-team suite in CI
  • Audit logs
  • Incident runbook

How an engagement usually runs

  1. Assessment2 weeks

    A threat model and red-team test of one AI system, with findings ranked by risk.

  2. Remediation4–8 weeks

    Controls built into the system, and a security test suite running in CI.

  3. Ongoing reviewMonthly

    Reviews of new AI features and regular re-testing.

Constraints we design for

Regulated data
Controls designed for environments under PIPA, DPDP and sector rules, with findings written for compliance teams.
Network separation (망분리)
Testing and tooling that run inside separated networks.
Third-party models and components
Provenance checks for model weights, packages and containers.
Existing security processes
Findings and fixes go through your vulnerability management and change process.

Typical stack

  • OWASP Top 10 for LLM Applications
  • MITRE ATLAS
  • garak
  • PyRIT
  • Open Policy Agent
  • Sigstore
  • Trivy
  • CycloneDX SBOM

We use the tools you already run where they fit. Names are for reference only.

Tell us about your systems

An engineer reads every message and replies within one business day.

Start a conversationHow we engage