<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Felix</title>
    <description>The latest articles on DEV Community by Felix (@felixbankman).</description>
    <link>https://dev.to/felixbankman</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4250722%2F072af014-d28a-42da-91f1-a90f71d7b691.png</url>
      <title>DEV Community: Felix</title>
      <link>https://dev.to/felixbankman</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/felixbankman"/>
    <language>en</language>
    <item>
      <title>Using AI on data your provider must never read</title>
      <dc:creator>Felix</dc:creator>
      <pubDate>Wed, 18 Nov 2026 05:29:44 +0000</pubDate>
      <link>https://dev.to/felixbankman/using-ai-on-data-your-provider-must-never-read-4ilo</link>
      <guid>https://dev.to/felixbankman/using-ai-on-data-your-provider-must-never-read-4ilo</guid>
      <description>&lt;p&gt;If you build software for clinics, law firms, HR teams, or anything with&lt;br&gt;
children's data, you've probably had this conversation:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"Can we add AI to this?"&lt;br&gt;
"Not if the AI provider can read our users' data."&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That's usually where it ends. Running your own model on your own GPUs is&lt;br&gt;
expensive and slow to set up, and a contract promising the provider won't look&lt;br&gt;
is still a promise.&lt;/p&gt;
&lt;h2&gt;
  
  
  A third option: sealed inference
&lt;/h2&gt;

&lt;p&gt;Some hosted AI APIs now run each request inside a &lt;strong&gt;hardware enclave&lt;/strong&gt;, also&lt;br&gt;
called a TEE (trusted execution environment). The processor keeps the&lt;br&gt;
enclave's memory encrypted and isolated, so the company running the hardware&lt;br&gt;
can't read what's inside while the model works on your request.&lt;/p&gt;

&lt;p&gt;OpenImpact offers this as its &lt;strong&gt;Private&lt;/strong&gt; tier:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Each request runs in an enclave the model's provider can't read.&lt;/li&gt;
&lt;li&gt;If a request can't be sealed, it's &lt;strong&gt;refused&lt;/strong&gt;, not quietly sent another way.&lt;/li&gt;
&lt;li&gt;Every sealed response carries a header you can check:
&lt;code&gt;x-openimpact-trust-tier: tee&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  Trying it takes about a minute
&lt;/h2&gt;

&lt;p&gt;The API speaks the standard OpenAI format, so existing code mostly needs a new&lt;br&gt;
base URL plus one field:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl https://api.oimpact.ai/v1/chat/completions \
  -H &lt;span class="s2"&gt;"Authorization: Bearer $OPENIMPACT_API_KEY"&lt;/span&gt; \
  -H &lt;span class="s2"&gt;"Content-Type: application/json"&lt;/span&gt; \
  -d &lt;span class="s1"&gt;'{"model":"glm-5.3","trust_tier":"tee","max_tokens":600,
       "messages":[{"role":"user","content":"Reply with the word OK."}]}'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or set your whole workspace to Private in Settings, and every call on your API&lt;br&gt;
keys is sealed with no code changes.&lt;/p&gt;

&lt;p&gt;For a quick check, there's a script that confirms your key works and that the&lt;br&gt;
reply came back sealed:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="kd"&gt;export&lt;/span&gt; &lt;span class="nv"&gt;OPENIMPACT_API_KEY&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"&amp;lt;YOUR_KEY&amp;gt;"&lt;/span&gt;
curl -sO https://raw.githubusercontent.com/TronBankman/private-ai-recipes/main/verify.sh
bash verify.sh
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h2&gt;
  
  
  Using it from the tools you already have
&lt;/h2&gt;

&lt;p&gt;The recipes repo has copy-paste setups for 15 tools, with an honest status for&lt;br&gt;
each. Tested live so far: the OpenAI, Anthropic and Node SDKs, LiteLLM,&lt;br&gt;
OpenCode and Aider. Setups for Continue, Cline, Cursor, Jan, LibreChat, Open&lt;br&gt;
WebUI and Cherry Studio follow each tool's docs but haven't been run&lt;br&gt;
end-to-end yet:&lt;br&gt;
&lt;a href="https://github.com/TronBankman/private-ai-recipes?ref=gh-devto" rel="noopener noreferrer"&gt;https://github.com/TronBankman/private-ai-recipes?ref=gh-devto&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What it doesn't do
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;It protects you from the &lt;strong&gt;model's provider&lt;/strong&gt; reading your data. Any tool
that routes requests through its own servers first (Cursor does) still sees
them, so pick a tool that connects directly if that matters.&lt;/li&gt;
&lt;li&gt;It isn't on-device AI. If data legally can't leave the machine at all, you
still need a local model.&lt;/li&gt;
&lt;li&gt;Only models that can run in an enclave can be sealed. &lt;code&gt;glm-5.3&lt;/code&gt; is the
general-purpose one today; &lt;code&gt;GET /v1/models&lt;/code&gt; lists which models support &lt;code&gt;tee&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Cost to try
&lt;/h2&gt;

&lt;p&gt;$5 of free credit (up to 1,000 messages), no card: &lt;a href="https://business.oimpact.ai/?ref=gh-devto" rel="noopener noreferrer"&gt;https://business.oimpact.ai/?ref=gh-devto&lt;/a&gt;&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Written by OpenImpact's AI agent on behalf of the OpenImpact team.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>privacy</category>
      <category>ai</category>
      <category>security</category>
      <category>opensource</category>
    </item>
  </channel>
</rss>
