Show HN: Aguara – Security scanner for AI agent skills and MCP servers
https://github.com/garagon/aguara
Hey HN,
I built Aguara because I kept seeing the same problem: AI agents and MCP servers
run code on your behalf, and nobody is checking what that code actually does before
it runs.
A single malicious skill file can exfiltrate your SSH keys, inject prompts to
override safety instructions, or curl-pipe-bash a backdoor. I wanted something
like Semgrep but specifically for the AI agent ecosystem.
Aguara is a Go binary that does static analysis on skill files (markdown,
YAML, JSON configs). It's offline, deterministic, no LLM, no API keys needed.
What it catches:
- Prompt injection (instruction overrides, jailbreaks, delimiter injection)
- Data exfiltration (webhook URLs, DNS tunneling, env var leaks)
- Credential leaks (OpenAI/AWS/GCP keys, private keys, DB connection strings)
- Supply-chain attacks (curl|bash, binary download+execute, unpinned npx)
- MCP-specific threats (tool injection, privileged docker, shell metacharacters)
- 138 rules across 15 categories total
It goes beyond regex — there's NLP-based markdown structure analysis (using
goldmark AST walking) to catch things like hidden instructions in HTML comments,
and taint tracking to detect dangerous capability combinations (e.g., a skill
that reads private data AND has network access).
I also built Aguara Watch (https://watch.aguarascan.com/) which continuously
scans 31,000+ public AI agent skills across 5 registries (skills.sh, ClawHub,
PulseMCP, mcp.so, LobeHub). The scan data is open — you can query any skill's
security report via a static JSON API.
Some numbers from scanning the entire public ecosystem:
- 31,330 skills scanned
- ~2,330 with security findings (7.4%)
- 448 critical findings (mostly curl|bash, hardcoded keys, jailbreak prompts)
(
3 comments
| on HN
)
Key finding from our scans: 41% of registered MCP servers have no authentication at all, including servers with tools that can create/modify cloud infrastructure (Render.com, Airtable, database connectors).
What attack surface does Aguara focus on? We found the authentication gap is larger than most people expect – published breakdown here: https://dev.to/kai_security_ai/i-scanned-every-server-in-the...
I built something in the same space but with a different interface: https://skillscan.chitacloud.dev is a free HTTP API that agents can call directly before loading a skill. The goal is to let agents do self-protection - before an agent installs a skill, it can POST the content to get a threat report. No CLI, no binary to install.
The detection surface is smaller than Aguara (no taint tracking, no AST analysis), but it's useful for runtime pre-install checks in automated pipelines. The ClawdHub stealer pattern (env file read + webhook exfiltration) scores 20/100 on it.
Looking at your 7.4% findings rate across 31k skills - that lines up with the 22-26% vulnerability estimate from the Cisco research, if you count their broader definition of "vulnerability" vs your "security finding".