Founding Security Cohort · 5 Design Partner Spots
Put your AI agents through real security tests.
Vigil is building runtime security and observability for AI agents, LLM applications, and MCP-connected workflows.
We're inviting up to five design partners to evaluate the product against realistic deployment patterns and help shape what runtime AI security should become.
Why this cohort exists
AI agents are moving from experimentation into real workflows. They interact with tools, retrieve external content, process documents, and make decisions that can affect real systems.
That creates security challenges that traditional application filters were not designed to handle.
We are building Vigil around these runtime risks and want to validate the product with teams facing them in practice.
Areas of evaluation
Evaluation focus
Prompt injection and untrusted context
Direct and indirect instructions arriving through user input, retrieved content, and external sources.
Tool and MCP security
Tool descriptions, schemas, tool outputs, unexpected behavior, and sensitive actions.
Obfuscation and multilingual input
Encoded content, Unicode manipulation, mixed-language prompts, and hidden instructions.
Policy enforcement and visibility
Security decisions, auditability, and the boundary between model suggestions and authorized application actions.
What design partners receive
- Guided technical onboarding and an initial threat-model session.
- A structured evaluation plan based on their deployment.
- A direct feedback channel with the Vigil product team.
- The opportunity to influence product priorities and integrations.
- A joint findings review, with publication only when both parties explicitly agree.
Who should apply
We are looking for teams that are:
- Building or operating LLM applications.
- Deploying AI agents or RAG systems.
- Connecting agents to tools or MCP servers.
- Responsible for application security, AI security, or platform engineering.
A technical owner and a controlled testing environment are expected.
Our evidence standard
We believe security claims should be backed by reproducible evidence.
We will not publish detection-rate, latency, certification, or universal language-coverage claims without supporting test data.
Design partner findings will remain private unless publication is explicitly approved.
Help shape the next layer of AI security.
Five design-partner spots. A focused working relationship. Real-world feedback.