AI agent security
See, control and defend every AI agent your clients run.
Copilot agents, coding assistants and MCP servers now read files, send email and run commands with real credentials. Phylax finds them, decides what each one may do before it acts, and flags it when it misbehaves. It is built for MSPs, universities and mid-sized security teams.
Core platform in development. Live demo opens within two weeks.
shell.exec("cat ~/.aws/credentials")Credential file read by a coding agentoutlook.send(to: "*@gmail.com", attach: payroll.xlsx)External send with sensitive attachmenttools/list changed: +delete_treeMCP server added a tool after it was approvedprompt contains SSN 6xx-xx-xxxxPersonal data leaving the endpointsharepoint.read("/Sales/Q3-plan.docx")Within the agent's allowed scopeThe problem
AI agents act with real access, and nobody can see them
The tools MSPs and security teams already run (RMM, EDR, email security) were never built to see what an agent actually does. Most companies cannot even say which employees run agents.
Copilot and Copilot Studio agents
Built by employees, not IT. They read SharePoint, send email and call connectors on the company's behalf.
Coding agents on every laptop
Claude Code, Cursor and GitHub Copilot run shell commands, read credential files and install packages.
New tools overnight
Any agent can gain tools in minutes. A server that quietly changes its tool list can turn a trusted tool hostile.
The platform
One console, three layers
Today an MSP would need three or four point tools to get this coverage. Phylax puts it in one multi-tenant console.
- InventoryEvery agent, MCP server, model server and identity, with owner and source.
- PostureMisconfigurations ranked by severity and environment.
- ExposureAttack paths and a 0–100 score with every point explained.
- GuardrailsAllow, deny, require approval or redact a tool call inline, before it runs.
- Agent identityWhich keys, service principals and OAuth grants each agent runs as.
- Detect and respondPrompt injection, data exfiltration and tool tampering, mapped to OWASP and MITRE ATLAS.
- AnalystAI triages and drafts; a person approves every decision.
- Board briefA plain-language report a CISO can forward upward.
Who it's for
Built for the people who secure everyone below the enterprise
Every engagement starts with a free 14-day discovery scan that shows each agent, MCP server and risky permission we find.
MSPs and MSSPs
One console for every client. Push a policy to 50 workspaces at once, and deploy through NinjaOne, ConnectWise, Datto, Kaseya, Intune or Jamf.
$5 per endpoint / month, wholesale
Universities and colleges
Title IV schools must meet the GLBA Safeguards Rule. Phylax covers the AI agents faculty, staff and students already use, with evidence for auditors.
From $12K per year, flat tiers
Mid-sized companies
In-house IT and security teams that are too small for enterprise vendors and too exposed to ignore AI agents. Defaults a small team can run.
$9 per endpoint / month, annual
How we earn trust
Safe to run inside your clients' environments
No LLM makes a blocking decision. A deterministic policy engine answers inline in milliseconds; AI detection runs in the background.
Detect mode first. Every guardrail starts by observing. Only a named admin turns on prevention.
Fail open by default. An outage on our side never breaks a client's agents.
Tenants are isolated. Database row-level security, with isolation tests on every change.
Deploys through your own RMM or MDM. Signed policies, with secrets kept on the device.
Honest coverage. Every report states exactly what is covered and since when.
Founder
Mark Castro
Founder and CEO · Austin, Texas · LinkedIn
- Security engineering at Meta and Google, including AI security.
- Started in the MSP world, so he knows what small and mid-sized clients need from their provider.
- Built the Phylax platform end to end: multi-tenant console, Go policy engine and Go endpoint sensor.
- A route into higher education through a UT Austin joint venture with another of his companies.
For investors
Our pre-seed round is open
Phylax by Vulnlogic is raising a pre-seed round on a SAFE, with angel checks welcome. This round takes the platform from build to paying clients:
- A production-ready product, live at a first pilot client
- Three or more paying clients, MSPs first
- A first university pilot
- A SOC 2 Type I report and a third-party pentest
Email us for the deck, the research behind our market sizing, and a live demo.
Contact
Demos, discovery scans and investor conversations
contact@vulnlogic.net
Copied
Design partners get a reduced pilot price, a direct line to the founder and real influence on the roadmap.