Cybersecurity Services for San Francisco Businesses

When AI Features Reach Customers

A feature built on a large language model (LLM) behaves differently from ordinary application code. It may read customer documents, follow instructions hidden inside them, or call tools on a user’s behalf. Conventional web scanners are not built to find these problems.

At the same time, enterprise buyers ask how customer data is protected and whether you are preparing for SOC 2. The services below address both pressures.

AI Application Security Testing

AI security testing can cover three layers, depending on what your product uses. LLM red teaming tries prompt injection, jailbreaks, data extraction, and output manipulation. Retrieval-augmented generation (RAG) reviews check who can retrieve which documents, whether the knowledge base can be poisoned, and where information leaks. Agent assessments test tool abuse, privilege escalation, limits on autonomous actions, and human approval steps.

Engagement Boundaries

Scope is built around your architecture, data sources, and user population, so testing targets the risks your design creates. Layers your product does not use are left out. Policy work, such as an AI governance framework, is included only when you scope it in.

What You Receive

Deliverables depend on the agreed scope and the systems assessed. Outputs can include an AI security assessment report and a prioritized remediation roadmap. Where LLM red teaming is in scope, you receive the findings with supporting evidence. A RAG assessment adds a security architecture review, and an agent assessment adds a controls blueprint.

LLM red team findings are mapped to the OWASP Top 10 for LLM Applications. That list of major LLM risks is maintained by the Open Worldwide Application Security Project (OWASP) community.

Cloud Security for SaaS Platforms

Your AI features run on cloud infrastructure that needs its own review. Cloud security engineering examines identity and access management (IAM), network segmentation, encryption, key management, and logging. Outputs can include an IAM strategy and federation blueprint, an encryption and key management plan, and a cloud security posture report.

SOC 2 Readiness and Implementation Support

SOC 2 examinations assess controls at service organizations. An independent certified public accountant (CPA) firm performs the examination and issues the report. SiegePal’s role is readiness and implementation support.

SOC 2 readiness starts with a gap assessment against the Trust Services Criteria. SiegePal then helps design and implement the missing controls, documents policies and procedures, and supplies evidence collection templates. Before the examination, you receive an audit preparation checklist and support answering auditor questions.

Other Core Services

The same practice offers penetration testing for the web application and application programming interface (API) around your AI features. If your platform handles protected health information, the Health Insurance Portability and Accountability Act (HIPAA) may be relevant, and HIPAA compliance work addresses it.

Relevant Project Experience

Our security portfolio and AI portfolio describe projects with their actual scope and status. The two below show relevant experience and do not imply San Francisco-based clients.

Ongoing security architecture review. SiegePal is currently reviewing the security architecture of a cloud-based AI infrastructure analytics platform. The work maps control gaps against SOC 2, HIPAA, and FedRAMP, the federal cloud authorization program. It also evaluates encryption and key management.

Completed AI engineering project. SiegePal built Aiperion’s generative AI engine using structured LLM output, including a custom Model Context Protocol (MCP) server and self-hosted speech-to-text. This was a build project, not a security assessment.

How an Engagement Runs

Scope the Architecture and Data

A free 30-minute consultation covers your AI features, cloud environment, data sources, and customer requirements. We identify which assessment activities fit the systems you actually use. The proposal defines the relevant deliverables and any readiness or implementation support.

Agree on the Testing Environment

Before AI testing begins, we agree on the environment, boundaries, and access required. Work is delivered remotely from a secure working environment. Connections can use a virtual private network (VPN) or another approved secure method suited to your requirements.

Review Findings and Next Steps

The assessment produces the outputs agreed for the systems in scope. Your team can address the findings, or SiegePal can support remediation when included in the engagement. For SOC 2 work, SiegePal supports readiness and implementation; an independent CPA firm performs the examination.

Questions San Francisco Buyers Ask

What should we check before an LLM feature goes live?

Assess it against the risks its design creates. That includes direct and indirect prompt injection, and downstream systems trusting model output without sanitizing it. It also includes retrieval access that is looser than the source data’s own permissions.

How will our developers understand AI findings?

Each LLM red team finding is mapped to an OWASP Top 10 for LLM Applications category. That gives your team a shared reference for understanding the issue and deciding what to fix first.

Can SiegePal help us write AI governance policies?

Yes, when it is in scope. An AI governance policy framework can cover responsible use guidelines, bias monitoring, and transparency requirements. It can also align with the National Institute of Standards and Technology (NIST) AI Risk Management Framework.