Add confidential AI examples

[h4x3rotab]

Summary

• Add confidential-ai/ directory with three examples:
  • inference - Private LLM inference with vLLM on Confidential GPU
  • training - Confidential fine-tuning using Unsloth
  • agents - Secure AI agent with TEE-derived wallet keys
• Update README with Confidential AI section

🤖 Generated with Claude Code

In general, the fix is to avoid returning raw exception messages to the client and instead return a generic error while logging the detailed error server-side. This preserves debuggability without exposing stack traces or potentially sensitive internal information to external users.

Concretely, in confidential-ai/agents/agent.py:

• For the / (index) route, change the except block so that it:
  • Logs the exception (with stack trace) using Python’s standard logging module.
  • Returns a generic JSON response such as {"status": "error", "message": "Internal server error"} with an appropriate HTTP status code (e.g., 500).
• To implement logging, add an import for logging at the top and optionally configure a basic logger (if not already configured elsewhere). Since we must not assume external configuration and can only use standard libraries, logging is appropriate and available.
• Do not change the normal (non-exceptional) behavior of the endpoint; only change what is returned when an exception occurs.

We will:

1. Add import logging near the other imports.
2. Replace the except Exception as e: body in index() so that:
   • It calls logging.exception("Error while handling index endpoint") (which logs the exception with a stack trace).
   • It returns a generic error JSON and a 500 status code, without including str(e).

No other routes need changes for this specific alert, as CodeQL only flagged line 144.

In general, the problem is that the /chat endpoint returns the string form of an arbitrary exception directly to the client, potentially exposing internal details. To fix this, the handler should return a generic, non‑sensitive error message to the user while optionally logging the detailed exception on the server for diagnostics.

The best minimal fix without changing functionality is: (1) introduce a logger for this module using Python’s standard logging library, and (2) in the /chat endpoint’s except block, log the exception (ideally with traceback) and replace {"error": str(e)} with a generic message like {"error": "Internal server error"}. This preserves the HTTP 500 status code and the basic contract that an "error" field is present, but avoids leaking specifics. We only need to modify confidential-ai/agents/agent.py: add a logging import and basic configuration near the top (or just get a logger if configuration is handled elsewhere), and update lines 172‑173 to log and send a generic response.

Concretely:

• At the top of agent.py, add import logging and a module‑level logger, e.g., logger = logging.getLogger(__name__). If the project doesn’t configure logging elsewhere, we can also add a simple basicConfig.
• In the /chat endpoint, change the except Exception as e: block to call logger.exception("Error while processing /chat request") (which logs the full traceback) and then return jsonify({"error": "Internal server error"}), 500 instead of using str(e).

To fix an unused-import issue, you remove the import statement for the module that is not referenced anywhere in the file. This reduces unnecessary dependencies and slightly simplifies the code without changing runtime behavior.

In this specific case, in confidential-ai/inference/verify.py, the json module imported on line 13 is not used anywhere in the shown code. The best fix is to delete that single import json line and leave the rest of the imports (argparse, hashlib, sys, requests) unchanged. No additional methods, definitions, or imports are required, and no other parts of the file need to be modified.

Concretely:

• Edit confidential-ai/inference/verify.py.
• Remove the line import json (currently line 13 in the snippet).
• Ensure all remaining imports and code remain exactly as they are.

To fix an unused import, the correct approach is simply to delete the import statement, as long as the imported module is not referenced anywhere in the file. This removes an unnecessary dependency and slightly improves readability and startup time.

For this file, confidential-ai/inference/verify.py, the single best fix is to remove the line import sys at line 14. No other code changes are needed, and no replacement functionality is required because the module was not used. Concretely: in verify.py, delete line 14 containing import sys while leaving all other imports and code unchanged.