Your AI Agents Are Installing
Dependencies.
Do You Know Which Ones?
AI coding assistants don't just write code — they install packages. Without a firewall that understands agent-level provenance, those installs happen outside your security policies. Hextrap is the only package firewall built natively for the agentic AI era.
Supply Chain Vulnerabilities now ranks in the top 3 LLM security risks. AI agents that autonomously install packages represent an attack surface that traditional package security tools — built before agents existed — cannot address.
OWASP LLM Top 10 →Traditional firewalls can't see agents.
Conventional package firewalls sit at the infrastructure layer. They know a package was requested — but they have no concept of who requested it. Was it a developer? A CI runner? An AI agent operating autonomously at 2am?
This matters because agents can install packages without developer review, based on instructions that may be malicious (prompt injection), ambiguous, or simply untested. A credential-harvesting package installed by your AI assistant looks identical to a legitimate install at the infrastructure layer.
In early 2026, compromised agent credentials were used to harvest data from 47 enterprise deployments for six months before detection. The vector: AI agents with unrestricted package installation capabilities.
- Agents install packages autonomously
- No distinction between human and agent installs
- Policy applies to CI/dev but not AI agents
- No audit trail of what the agent installed
- Prompt injection can bypass package restrictions
- Every agent install passes through your firewall
- Agent identity logged alongside every request
- Same allow/deny policies apply to AI and humans
- Full audit log: agent → package → policy result
- Agents can only install what your policy permits
How the MCP integration works
Hextrap exposes a Model Context Protocol server. Your AI assistant connects to it and uses it as a tool when it needs to install packages.
Developer asks AI assistant
"Install the latest version of httpx and add it to requirements.txt"
AI agent calls Hextrap MCP tool
The agent invokes hextrap_check_package before installing
Hextrap evaluates the request
Checks typosquats, soak time, allow/deny lists, and threat intelligence
Approved or blocked — with reason
Agent proceeds only if cleared. Result logged with agent identity and timestamp.
Full visibility into what your AI is doing.
Every package install by an AI agent is logged separately from developer and CI installs. You can see exactly what your AI assistant has installed, when, and whether it passed your policies.
- Agent identity captured per request
- Package, version, and registry logged
- Policy result (allowed/blocked/quarantined) and reason
- Timestamp and session context
- Filterable by actor type: developer, CI, AI agent
This isn't just useful for security — it's the audit trail your compliance team will ask for when AI governance frameworks start requiring documentation of autonomous system behavior.
Works with the tools your team already uses.
Any AI coding assistant that supports Model Context Protocol can connect to Hextrap.
Govern your AI agents before your security team asks you to.
Five minutes to set up. Zero workflow changes for your developers.