diff --git a/.github/workflows/trojanchat-hygiene.yml b/.github/workflows/trojanchat-hygiene.yml
index df06253..655a400 100644
--- a/.github/workflows/trojanchat-hygiene.yml
+++ b/.github/workflows/trojanchat-hygiene.yml
@@ -112,6 +112,7 @@ jobs:
         run: |
           pip-audit -r requirements.txt --progress-spinner off --format json --output pip-audit.json
       - name: Upload dependency audit
+        if: always()
         uses: actions/upload-artifact@v7
         with:
           name: pip-audit-report
diff --git a/Dockerfile b/Dockerfile
index 13a1919..d7b0552 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -18,10 +18,11 @@ ENV PATH=/opt/venv/bin:$PATH \
     PYTHONDONTWRITEBYTECODE=1 \
     PYTHONPATH=/app
 
-# The base image ships build tooling that is unnecessary at runtime. Removing it
-# also prevents scanners from treating setuptools' vendored build dependencies
-# as remotely exploitable application packages.
-RUN python -m pip uninstall --yes setuptools wheel \
+# The runtime executes Uvicorn from /opt/venv and does not install packages.
+# Remove the base image's package-management tooling and bundled installer
+# payloads so they cannot ship vulnerable vendored dependencies.
+RUN python -m pip uninstall --yes pip setuptools wheel \
+    && rm -rf /usr/local/lib/python3.14/ensurepip \
     && useradd --create-home --uid 10001 appuser
 WORKDIR /app
 
diff --git a/ml/experiments/prompt_eval.py b/ml/experiments/prompt_eval.py
deleted file mode 100644
index 3e0402c..0000000
--- a/ml/experiments/prompt_eval.py
+++ /dev/null
@@ -1,12 +0,0 @@
-import time
-from app.core.llm_client import LLMClient
-from ml.tracking.mlflow_utils import log_chat_metrics
-
-def run_experiment(prompt: str):
-    llm = LLMClient(api_key="DUMMY")
-    start = time.time()
-    output = llm.generate(prompt)
-    latency = (time.time() - start) * 1000
-
-    log_chat_metrics(latency, len(output.split()), "v1")
-    return output
\ No newline at end of file
diff --git a/ml/ml/experiments/prompt_eval.py b/ml/ml/experiments/prompt_eval.py
deleted file mode 100644
index 2fccc04..0000000
--- a/ml/ml/experiments/prompt_eval.py
+++ /dev/null
@@ -1,19 +0,0 @@
-import time
-from app.core.llm_client import LLMClient
-from ml.tracking.mlflow_utils import start_run, log_metrics, log_params
-from ml.metrics import token_count
-
-def run_prompt_experiment(prompt: str, api_key: str):
-    llm = LLMClient(api_key)
-    with start_run("prompt-eval-v1"):
-        start = time.time()
-        output = llm.generate(prompt)
-        latency = (time.time() - start) * 1000
-
-        log_params({
-            "model": "gpt-4.1",
-            "prompt_version": "v1"
-        })
-        log_metrics(latency, token_count(output))
-
-    return output
\ No newline at end of file
diff --git a/ml/tracking/mlflow_utils.py b/ml/tracking/mlflow_utils.py
deleted file mode 100644
index 0a41b15..0000000
--- a/ml/tracking/mlflow_utils.py
+++ /dev/null
@@ -1,13 +0,0 @@
-import mlflow
-
-def start_run(run_name: str):
-    mlflow.set_experiment("trojanchat-llm")
-    return mlflow.start_run(run_name=run_name)
-
-def log_metrics(latency_ms: float, tokens: int):
-    mlflow.log_metric("latency_ms", latency_ms)
-    mlflow.log_metric("tokens_used", tokens)
-
-def log_params(params: dict):
-    for key, value in params.items():
-        mlflow.log_param(key, value)
\ No newline at end of file
diff --git a/requirements.txt b/requirements.txt
index 8faf598..2d8783a 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,7 +1,6 @@
 openai>=1.68.0
 groq>=1.1.2
 streamlit>=1.44.0
-mlflow>=2.22.0
 pydantic>=2.10.0
 pydantic-settings>=2.8.0
 python-dotenv>=1.1.0
@@ -14,3 +13,8 @@ prometheus-client>=0.21.0
 PyJWT>=2.12.0
 passlib[bcrypt]>=1.7.4
 python-json-logger>=2.0.7
+
+# Security floors for transitive packages identified by the container scan.
+cryptography>=50.0.0
+msgpack>=1.2.1
+setuptools>=78.1.1