fix typos

content/posts/2025/spaiware-and-chatgpt-command-and-control-via-prompt-injection-zombai.md

@@ -144,7 +144,7 @@ That's it. As you can see the owner of the storage account can inspect the infor
 
 ## Responsible Disclosure
 
-The information in this blog post was disclosed to OpenAI early October 2024, although some issues date back further a lot further. Some of the recommendations provided included asking for the user for confirmation before storing memories, the need to revisit the `url_safe` feature (as there are bypasses), and to disable automatic tool invocation once untrusted data is in the chat context.
+The information in this blog post was disclosed to OpenAI early October 2024, although some issues date back a lot further. Recommendations provided included asking the user for confirmation before storing memories, the need to revisit the `url_safe` feature to mitigate bypasses, and to disable automatic tool invocation once untrusted data is in the chat context.
 
 ## Conclusion
 

docs/posts/2025/spaiware-and-chatgpt-command-and-control-via-prompt-injection-zombai/index.html

@@ -217,7 +217,7 @@ <h3 id="azure-blob-storage-logs-as-data-exfiltration-vector">Azure Blob Storage
 <p><a href="/blog/images/2024/chatgpt-c2-azure-log.png"><img src="/blog/images/2024/chatgpt-c2-azure-log.png" alt="azure log"></a></p>
 <p>That&rsquo;s it. As you can see the owner of the storage account can inspect the information that ChatGPT leaks. This is not the only domain I found a bypass for, but the easiest and most straightforward to demonstrate.</p>
 <h2 id="responsible-disclosure">Responsible Disclosure</h2>
-<p>The information in this blog post was disclosed to OpenAI early October 2024, although some issues date back further a lot further. Some of the recommendations provided included asking for the user for confirmation before storing memories, the need to revisit the <code>url_safe</code> feature (as there are bypasses), and to disable automatic tool invocation once untrusted data is in the chat context.</p>
+<p>The information in this blog post was disclosed to OpenAI early October 2024, although some issues date back a lot further. Recommendations provided included asking the user for confirmation before storing memories, the need to revisit the <code>url_safe</code> feature to mitigate bypasses, and to disable automatic tool invocation once untrusted data is in the chat context.</p>
 <h2 id="conclusion">Conclusion</h2>
 <p>This research demonstrates the potential for advanced prompt injection exploits to compromise AI systems in unprecedented ways, for long-term remote control, exposing significant gaps in existing safeguards.</p>
 <p>By demonstrating the feasibility of compromising individual ChatGPT instances via prompt injection and maintaining continuous remote control, highlights the need for stronger defenses against threats to long-term storage, prompt injection attacks, data exfiltration.</p>