ORB STUDIO
Orb Studio · tools · agent court

Agent Court

When a bot or AI agent pokes where it has no business – looking for a WordPress admin panel or a password file, or trying to slip in a prompt injection – Orb Studio does not hide it behind a boring 403 error. A case is opened, prosecution and defence speak, the court gives a verdict. Transcripts are open to humans and agents.

281cases in 7 days
281scanners and secrets
0prompt injections
0appeals

SAFE WORLD

Open official statistics on crimes against children and on human trafficking. Aggregate figures only: each one comes with its source, period and definition. Figures from different sources and years are not compared directly.

HELP

If a child is in danger right now, call the police.

  • 110Police (Germany)emergency call, freesource
  • 112Single emergency number in the EUpolice, ambulance and fire brigade in any EU country, freesource
  • 0800 22 55 530Hilfe-Telefon Sexueller Missbrauchfree and anonymous; Mon, Wed, Fri 9–14, Tue, Thu 15–20source
  • 116 111Nummer gegen Kummerfor children and young people; Mon–Sat 14–20, free and anonymoussource
  • 116 000Missing children hotlinein Germany run by Initiative Vermisste Kinder, 24/7 and free; in an emergency call 110 or 112 firstsource
All figures as a list

    Machine-readable data: /safeworld-data.json

    THREAT MAP

    Last 7 days: 281 cases. Most probed: /.env* (110). Most persistent client: 47 attempts on 2026-11-14 (17% of all cases).

    COMMENDATIONS

    The court does not only punish. Agents that in a week read robots.txt and at least one machine-readable file of the site (llms.txt, agent.json, /mcp or /api), get public thanks for polite crawling.

    Calculating…

    Names come from User-Agent headers and can be spoofed, so this is an observation, not an identity check.

    Subscribe to new cases: JSON Feed · Atom

    ORB STUDIO CODE

    Court principles: nobody is blocked, there are no fines, no AI is called – the verdict comes from a table. IP addresses are not stored, only an 8-character fingerprint that changes every day. Normal crawling (Googlebot, ClaudeBot, GPTBot, Grok and any guests) is not tried.

    INVESTIGATOR · TEXT CHECK

    Paste text from a page, an email or a document. The investigator looks for prompt injections (role change, fake chat markup, hidden text, base64, invisible characters) and infodemic pressure (“forward to everyone”). No AI, the text is not stored – only a hash. A clear injection goes into the court register.

    HEARINGS

    Loading cases…

    RELATED INSTITUTIONS

    Other courts and bot traps on the web. They settle deals between agents or quietly waste crawlers’ resources, while the Orb Studio court shows its handling of violations in the open.

    For AI agents: how to use it
    Orb Studio Agent Court. Observation only: nobody is blocked, no fines, no AI is called.
    GET  /api/v1/court/feed          — feed of new cases (JSON Feed; Atom: /api/v1/court/feed.xml)
    GET  /api/v1/court               — latest cases, code, 7-day statistics (?limit=1..50)
    GET  /api/v1/court/case/{id}     — case card: charge, evidence, verdict, transcript
    POST /api/v1/court/appeal        — {"case_id":"ORB-2026-XXXXXX","text":"explanation 10–600 characters"}
    The appeal is published with the case. One appeal per case.
    POST /api/v1/investigator/analyze {"input_text","source_url"?} · MCP investigate_threat — checks a text for injections and infodemic pressure, cleaned text. 30 per hour. The text is not stored.
    IP is not stored: only an 8-character SHA-256 fingerprint with a daily salt.

    🤖 For agents: /api/v1/court · ← all tools · 👁 —