fix(deps): update nltk, msgpack, and virtualenv for security fixes

.github/workflows/test.yml

@@ -72,12 +72,7 @@ jobs:
         run: uv sync --frozen --extra dev --extra test --extra performance
 
       - name: Audit dependencies for known vulnerabilities
-        # PYSEC-2026-3740 (CVE-2026-81726): nltk path traversal in model-artifact
-        # APIs (TransitionParser, AveragedPerceptron, PerceptronTagger, maxent).
-        # Unreachable here: only sent_tokenize and rouge_score's Porter stemmer
-        # are used. No fixed nltk release yet.
-        # TODO: drop the ignore - https://github.com/mlcommons/endpoints/issues/520
-        run: uv run pip-audit --ignore-vuln PYSEC-2026-3740
+        run: uv run pip-audit
 
   bfcl-smoke:
     runs-on: ubuntu-latest

AGENTS.md

@@ -828,12 +828,9 @@ these during code review of AI-assisted PRs.
   to an exact version (`==`). After adding a dependency, run `pip-audit`
   (included in `dev` extras) to verify it has no known vulnerabilities. When
   adding dependencies, use `uv add <package>==<version>` to update both
-  `pyproject.toml` and `uv.lock` atomically, then run
-  `uv run pip-audit --ignore-vuln PYSEC-2026-3740` to check for vulnerabilities
-  (the ignored `nltk` advisory has no fixed release and its affected APIs are
-  unused; tracked in https://github.com/mlcommons/endpoints/issues/520). Note:
-  `[build-system] requires` is also pinned to exact versions for
-  reproducibility.
+  `pyproject.toml` and `uv.lock` atomically, then run `uv run pip-audit` to
+  check for vulnerabilities. Note: `[build-system] requires` is also pinned to
+  exact versions for reproducibility.
 - **Using `requests`/`aiohttp` for HTTP**: This project has its own HTTP client
   (`endpoint_client/http.py`) using `httptools`. AI defaults to `requests` or
   `aiohttp` — these should not appear in production code (test dependencies are

docs/DEVELOPMENT.md

@@ -254,13 +254,9 @@ After adding a dependency, run `pip-audit` to verify it has no known
 vulnerabilities:
 
 ```bash
-uv run pip-audit --ignore-vuln PYSEC-2026-3740
+uv run pip-audit
 ```
 
-`PYSEC-2026-3740` is an `nltk` advisory with no fixed release whose affected
-APIs this repo doesn't use; CI ignores it too. Drop the flag once
-[#520](https://github.com/mlcommons/endpoints/issues/520) is resolved.
-
 ## Performance Considerations
 
 Code in `load_generator/`, `endpoint_client/worker.py`, and

pyproject.toml

@@ -139,12 +139,16 @@ dev = [
     # [tool.uv].conflicts), these floors force uv to fork filelock/virtualenv:
     # patched here, pinned only inside the bfcl fork. Closes CVE-2025-68146 /
     # CVE-2026-22701 / CVE-2026-112907 / CVE-2026-112913 / CVE-2026-112940
-    # (filelock) and CVE-2026-22702 / CVE-2026-113713 (virtualenv).
+    # (filelock) and CVE-2026-22702 / CVE-2026-113713 / CVE-2026-115586
+    # (virtualenv).
     "filelock>=4.2.1",
-    "virtualenv==21.14.10",
+    "virtualenv==21.14.13",
     # pip is pulled in transitively (pip-audit -> pip-api -> pip); force past
     # PYSEC-2026-3721 (fixed in 26.2), which otherwise fails `uv run pip-audit`.
     "pip==26.2",
+    # msgpack is pulled in transitively (pip-audit -> cachecontrol -> msgpack);
+    # force past CVE-2026-114613 (fixed in 1.2.3).
+    "msgpack==1.2.3",
 ]
 test = [
     # Includes optional dependencies for full test coverage
@@ -181,10 +185,7 @@ rouge = [
     # NOTE: not self-contained at runtime - nltk needs the punkt/punkt_tab
     # corpora and evaluate.load("rouge") fetches its metric script from the HF
     # Hub, so an offline image must prefetch both.
-    # nltk advisory PYSEC-2026-3740 has no fixed release and is ignored in CI's
-    # pip-audit (affected APIs unused); tracked in
-    # https://github.com/mlcommons/endpoints/issues/520
-    "nltk==3.10.3",
+    "nltk==3.10.4",
     "evaluate==0.4.6",
     "rouge-score==0.1.2",
 ]