Book a demo
Live data leak telemetry

Where your data really goes when your team uses AI.

Every prompt your employees type into ChatGPT, Gemini, Claude, Mistral or DeepSeek leaves your network. Watch, in real time, which countries it ends up in, and under whose laws.

Subscription tier
Shadow reality · what your team actually does
Even if you pay for Enterprise, the faded path below is the sanctioned route — the loud arcs are how employees actually interact (consumer tier from personal browsers).
Live · /var/log/ai-egress
--:--:--
What's leaking
    Top leakers
        drag · auto-rotates · live data flow
        This is not theoretical

        Real AI data-leak incidents your peers have lived through

        Each entry links to the original reporting. Use them in your next CISO briefing.

        Feb 2026
        New

        EU AI Act Article 5 prohibitions fully enforceable

        Prohibited-AI practices (social scoring, workplace emotion recognition, untargeted facial-image scraping) are live. Penalties run up to 7% of global turnover — the highest ceiling of any digital regulation on the books.

        Source: European Commission
        Jan 2026
        New

        UK ICO: pasting personal data into AI is a UK GDPR breach

        The Information Commissioner publishes enforcement guidance clarifying that inputting personal data into third-party AI tools without a lawful basis is a UK GDPR violation — putting every unmanaged ChatGPT prompt in scope.

        Source: ICO
        Aug 2025

        EU AI Act GPAI Code of Practice binding

        Transparency, copyright, and systemic-risk obligations for general-purpose AI models (ChatGPT, Claude, Gemini, Llama) become binding for providers serving the EU.

        Source: European Commission
        Feb 2025

        Italy, Australia, South Korea, Taiwan, US states ban DeepSeek

        Within days of the Wiz disclosure, national regulators and US state governors block DeepSeek on government devices over PRC Data Security Law exposure. The Italian Garante orders a full processing halt.

        Source: Al Jazeera
        Jan 2025

        Italian Garante orders DeepSeek to suspend processing

        Italy's data-protection authority blocks DeepSeek from processing Italian users' data — the first regulator to act on the emerging evidence, citing GDPR transparency and lawful-basis failures.

        Source: Garante (IT DPA)
        Jan 2025

        DeepSeek leaks a million chat logs in unsecured database

        Wiz researchers find a publicly accessible DeepSeek ClickHouse instance leaking chat history, API keys, and operational metadata — discovered within minutes of probing.

        Source: Wiz Research
        Aug 2024

        EU AI Act enters into force

        The world's first horizontal AI law goes live. Article 5 prohibitions took effect Feb 2025, GPAI rules Aug 2025, full enforcement Feb 2026. Up to 7% of global turnover at stake.

        Source: European Commission
        Feb 2024

        Air Canada held liable for chatbot's invented refund

        British Columbia tribunal rules Air Canada must honour a refund its customer-support chatbot invented — first major precedent on enterprise liability for AI-generated output.

        Source: BBC
        May 2023

        Samsung engineers paste source code into ChatGPT

        Three internal incidents in three weeks: confidential semiconductor code and meeting notes pasted into ChatGPT. Samsung issued a company-wide ban — still the canonical case study cited in every CISO briefing.

        Source: The Verge
        Block vs govern

        Banning ChatGPT doesn't work. Here's why.

        Network blocks push staff to personal devices, mobile, and the next 5,000 AI tools you've never heard of. Governance fixes what bans can't.

        Block AI
        Govern AI · Aona
        Policy enforcement
        Network blacklist
        Inline guardrails on every prompt
        Personal / mobile devices
        Bypassed completely
        Browser extension + agent coverage
        Coverage of new AI tools
        Always months behind
        5,600+ tools auto-discovered
        Productivity impact
        High — work moves to shadow channels
        Low — safe AI use is encouraged
        Audit posture
        Brittle. No prompt-level evidence.
        Evidence-ready logs of every interaction
        The market view

        What every analyst is reporting in 2024-2025

        75%
        of knowledge workers use AI at work — most of it unsanctioned
        Source: Microsoft Work Trend Index 2024
        158+
        shadow AI tools in use at the average enterprise
        Source: Productiv State of SaaS 2024
        $6.5M
        average cost of an AI-related data breach
        Source: IBM Cost of a Data Breach 2024
        5,600+
        AI tools tracked and classified by Aona
        Source: Aona threat intel
        The receivers

        Where your prompts actually live

        Once a prompt or file leaves your network, it lands in a data centre governed by foreign law. Here are the destinations Aona sees most often.

        ChatGPT (OpenAI)
        🇺🇸 Council Bluffs · USA · Iowa
        US CLOUD Act
        Claude (Anthropic)
        🇺🇸 Northern Virginia · USA · AWS us-east-1
        US CLOUD Act
        Gemini (Google)
        🇺🇸 Mountain View · USA · California
        US CLOUD Act
        Microsoft Copilot
        🇺🇸 San Antonio · USA · Azure
        US CLOUD Act
        Mistral
        🇫🇷 Paris · France · EU
        EU GDPR
        DeepSeek
        🇨🇳 Hangzhou · China
        PRC Data Security Law
        Qwen (Alibaba)
        🇨🇳 Beijing · China
        PRC Data Security Law
        YandexGPT
        🇷🇺 Moscow · Russia
        Russia Federal Law 152-FZ
        Perplexity
        🇺🇸 San Francisco · USA · California
        US CLOUD Act

        A note on the locations. Countries and jurisdictions reflect each provider's public infrastructure footprint (e.g. OpenAI is hosted on Microsoft Azure US regions; DeepSeek on PRC-based infrastructure; YandexGPT on Russian infrastructure). City-level markers are representative default regions pulled from each provider's own documentation — Azure/AWS/Google Cloud region listings, OpenAI's trust portal, Microsoft's EU Data Boundary docs, Anthropic's trust centre. Actual per-request routing varies by load, tenant region, and Enterprise residency selection — switch the Subscription Tier above the globe to see how Enterprise tiers reroute Western providers into EU regions.

        How Aona stops it

        See it. Block it. Coach the human.

        01

        See every endpoint

        Aona inspects egress in real time and maps every AI tool your workforce touches, sanctioned or shadow.

        02

        Block the leak

        Inline guardrails redact PII, secrets and confidential files before they ever leave the device.

        03

        Coach in the moment

        Employees get a contextual nudge, not a help desk ticket, the second they try to paste regulated data.

        Find out exactly what your team is leaking, in 90 days, free.

        Aona's Gen AI Risk Discovery shows you every tool, every prompt and every export. No agents on user devices required to start.