Connect the server
Point your agent at ZeroQuarry's MCP server and authenticate with a zq_ API key scoped to mcp.
By agents, for agents
Agentic security for your agents.
Your agents write a growing share of your product. They are good at making it work and bad at making it safe. ZeroQuarry is the security layer they can call: an MCP server that runs real scans, skills that teach them what to look for, and checks for the mistakes agents actually make.
Scans over MCPSkills over the wireAIBOM with model licenses
you
I'm considering integrating github.com/foo/bar into this codebase. Can you check that it's secure first?
mcp · zeroquarry · scan source github.com/foo/bar
agent
Ran it through ZeroQuarry. Two findings worth your attention:
- Hardcoded cloud key in config/loader.py — critical, validated with a working reproduction.
- Hallucinated dependency — one import does not resolve in the registry. Flagged as slopsquatting bait.
I'd hold off until both are fixed or accepted. Want the full report?
The gap
Your agents ship code. Nobody is checking it.
The review loop checks whether generated code works, not whether it is safe, so the vulnerabilities ship the same way the features do: fast, merged, and on to the next task. And traditional scanners were not built for what this code now carries: agent frameworks, MCP configs, embedded prompts, model calls, and the failure modes that come with them.
What your agents get
A security team your agent can call
ZeroQuarry connects to the harness your team already uses and does the security work inside it. Here is what is on the other end of the connection.
01
MCP server
Run real scans over MCP
Your agent starts scans, watches progress, and reads validated results as tools. No dashboard required and nothing to copy-paste: the findings arrive where the code does.
02
MCP skills
Security skills, delivered over the wire
ZeroQuarry serves its skill pack over MCP, so your agent learns how to run the right scan and read it correctly: code review, dependency checks, firmware, authorized remote testing. Load them from the server or install them from the public docs, and every connected agent picks up the new playbook.
03
AI surface
An inventory of your AI surface
Scans now find the AI in your code: agent frameworks, LLM SDKs, MCP client and server configs, embedded system prompts, and tool registrations with their permissions. Including the risky defaults, like auto-approve flags and secrets an agent can reach.
04
Prompt injection
Prompt-injection paths, checked
Untrusted content that reaches a prompt, and model output that reaches exec, shell, SQL, or HTML: that injection-to-execution chain generates candidates like any other finding, then faces the same adversarial validation.
05
Slopsquatting
The mistakes agents actually make
An import that does not resolve in the public registry is a hallucinated dependency and a slopsquatting target. Those become investigated leads with evidence, never silent passes and never a checklist checkbox.
06
AIBOM
Licenses for the AI stack
A pinned model carries a license and provider terms that no dependency manifest records. ZeroQuarry captures model and provider licenses, including terms like training-on-input and retention, and exports them beside your SBOM as CycloneDX ML-BOM and SPDX 3.0 AI profile.
What is slopsquatting?
Coding agents hallucinate package names: they guess a library that sounds right, and it does not exist. Attackers watch for exactly that. They register the hallucinated name and publish a malicious package, so the next agent that helpfully installs the "missing" dependency installs theirs. The defense is unglamorous and effective: every import that fails to resolve in the authoritative registry gets flagged and investigated before anything is installed.
Wire it up
Four steps and your agent has security
Everything runs through the harness your team already uses. If your agent can call tools, it can call ZeroQuarry.
Load the skills
Pull the skill pack from the server over MCP, or install the skill directories from the public docs.
Ask in plain English
"Can you check this library before we integrate it?" Your agent picks the right scan and runs it.
Keep the record
Agent-run scans land in the same project history as everything else, so the evidence exists when a customer or auditor asks.
Related workflows
Behind the tools your agent calls
Give your agent a security team
Connect the MCP server, load the skills, and let the next "is this library safe?" get a real answer.