Every listing is security-scanned, automatically — before anyone installs it
Every listing is security-scanned, automatically — before anyone installs it
AI capabilities run inside real systems with real permissions. An MCP server can read files. A guardrail can intercept model output. A dataset can embed adversarial examples. The risk surface is real — and we scan for all of it.
Every artifact uploaded to ai-supply.store is automatically scanned before it becomes installable. This is non-negotiable, non-bypassable, and free.
What the scanner checks
Nine distinct check layers run in parallel on every upload:
| Layer | What it catches |
|---|---|
| Malware | Known payloads, shellcode, obfuscated scripts |
| Secrets | Hardcoded API keys, tokens, private keys |
| Dangerous code | eval() abuse, shell injection, network backdoors |
| PII | Emails, phone numbers, SSNs in datasets and prompts |
| License | GPL/AGPL contamination, missing attribution |
| Dependency CVEs | Vulnerabilities in package.json, requirements.txt, lockfiles |
| Model format | Pickle exploits, malformed GGUF/safetensors, hidden execution |
| Prompt injection | Instructions designed to hijack downstream agent behaviour |
| Egress | Unexpected outbound network call patterns |
OWASP-AI on every listing
Beyond the automated layers, every listing is evaluated against the full OWASP AI Security Top 10 (LLM01–LLM10) and the OWASP ML Security Top 10 (ML01–ML10). The results appear as an expandable checklist on the Security tab of every listing page.
Score, grade, level
After scanning, every listing gets:
- A score from 0 to 100
- A grade of A, B, C, or D
- A safety level:
SAFE,REVIEW, orQUARANTINE
Critical findings → QUARANTINE. The listing is blocked. The provider must fix and re-submit. There is no override.
Grade A listings appear on the Most secure leaderboard. That leaderboard is updated in real time as new versions are scanned.
Deep scan engines
When configured, the scanner also invokes:
- Opengrep — AST-level code analysis and taint tracking
- picklescan — model-format malware detection for pickle-based artifacts
- Gitleaks — deep secrets scanning across the full artifact tree
- osv-scanner — CVE lookups against the Open Source Vulnerability database
Every new version upload triggers a fresh scan. Security posture is tracked over time.
Read the full technical breakdown in the nine-layer scanner: a deep dive.