OSINT: detect exposed AI infrastructure (and don’t invent findings)

Ship in /var/Agentesosint_ai_exposure, origin IP behind CDN/WAF, façade of the 5k-line monster, coverage of CLEAN tools in the scan report, and SSRF/framing fixes in multi-agent review.

What “exposed AI” means

Open inference endpoints, unauthenticated panels, keys in JS, public vector stores, accessible LLM consoles. The Code Orchestrator and the security workflow include it in the smart scan — not as a “generic security” report but as an AI surface.

Report quality

• No filler: if theHarvester returns empty, it says empty (with reason), nothing is invented.
• Subdomains unified in one engine
• Fallback ladder in search: the expensive provider comes last
• UI Design review with headless Chromium: mobile overflow and real contrast

Cloudflare joins the MCP marketplace: docs without auth + Code Mode with token when needed.

OSINT audit of your domain

What we look for when we talk about “exposed AI”

It is not a full pentest: it is a map of the public attack surface where inference panels, unauthenticated APIs, buckets, staging subdomains and MCP documentation often slip through. ADRI's OSINT agent combines search, scrapers and classic tools (e.g. email/subdomain harvesting) and falls back from a cheap to an expensive provider only when needed.

• Subdomains unified in one engine (less noise, fewer duplicates).
• If a tool returns empty, the report says so — it does not invent CVEs.
• Optional UI review with headless Chromium when the finding is a web console.

Ethical and legal limits

OSINT is limited to open sources and targets that the client authorises. It is not a free pass to attack third parties. The pentester agent issues findings; it does not “fix” on the fly without a mandate.

Frequently asked questions

More: OSINT and pentester agents · security with agents · domain audit.

• OSINT
• Security
• AI Exposure
• Code Orchestrator