fix(cpex): bump litellm to 1.84.0 to clear dependency-review vulns

litellm 1.83.10 (the previous pin) is flagged by two newer advisories that
fail the PR dependency-review gate (fail-on-severity: moderate):

  * GHSA-4xpc-pv4p-pm3w — authentication bypass via Host header injection
    (critical; vulnerable < 1.84.0)
  * GHSA-qrc4-49gv-mv9m — internal_user can create API keys beyond their
    role (high; vulnerable < 1.83.14)

1.84.0 is the lowest release that clears both (and every other known
litellm advisory), consistent with the file's lowest-safe-release pin
convention.

Assisted-By: Claude (Anthropic AI) <noreply@anthropic.com>
Signed-off-by: Frederico Araujo <frederico.araujo@ibm.com>

Author: araujof

authbridge/demos/hr-cpex/agent/requirements.txt

@@ -17,11 +17,14 @@ uvicorn[standard]==0.32.0
 # LLM provider abstraction — supports OpenAI, Anthropic, Ollama, WatsonX,
 # Azure, Bedrock, and more, so the demo runs against whatever the operator
 # has access to without code changes. litellm is pinned at the lowest safe
-# release: 1.82.7 / 1.82.8 had a supply-chain compromise (1.83.0 floor) and
+# release: 1.82.7 / 1.82.8 had a supply-chain compromise (1.83.0 floor),
 # 1.83.0–1.83.9 carry a SQL-injection / SSTI / RCE / sandbox-escape cluster
-# (GHSA-r75f-5x8p-qvmc, -xqmj-j6mv-4862, -v4p8-mg3p-g94g, -wxxx-gvqv-xp7p),
-# all patched by 1.83.10.
-litellm==1.83.10
+# (GHSA-r75f-5x8p-qvmc, -xqmj-j6mv-4862, -v4p8-mg3p-g94g, -wxxx-gvqv-xp7p,
+# patched by 1.83.10), and < 1.84.0 carries an authentication bypass via
+# Host header injection (GHSA-4xpc-pv4p-pm3w, critical) plus an
+# internal_user API-key privilege escalation (GHSA-qrc4-49gv-mv9m, high) —
+# both patched by 1.84.0.
+litellm==1.84.0
 
 # httpx for the MCP tool call (routed through the sidecar forward proxy).
 # Bumped to 0.28.1 to satisfy a2a-sdk's httpx>=0.28.1 floor.