critical

CVE

CVE-2026-116692, CVE-2026-116690, CVE-2026-116691, CVE-2026-116684, CVE-2026-116685, CVE-2026-116686, CVE-2026-115419, CVE-2026-115413, CVE-2026-115522, CVE-2026-115883, CVE-2026-118668, CVE-2026-118667, CVE-2026-115885, CVE-2026-116575, CVE-2026-115886, CVE-2026-118623, CVE-2026-118627, CVE-2026-118628

CWE

CWE-94, CWE-22, CWE-59, CWE-532, CWE-488, CWE-668, CWE-706, CWE-862, CWE-346, CWE-918, CWE-502

Affected Surface

  • PyPI `mlflow` versions `>= 3.12.0rc0, < 3.17.0` (`CVE-2026-116692` CVSS 9.8 RCE and `CVE-2026-115886` SSRF), `>= 3.10.0, < 3.17.0` (`CVE-2026-116690` arbitrary file read), and `>= 2.19.0, < 3.17.0` (`CVE-2026-116691` CrewAI autolog API-key disclosure), plus ZDI advisories `CVE-2026-118623` (`ZDI-26-839`), `CVE-2026-118627` (`ZDI-26-843`), and `CVE-2026-118628` (`ZDI-26-844`)
  • PyPI `mcp` (Model Context Protocol Python SDK) versions `>= 1.8.0, < 1.30.0` and `>= 2.0.0, < 2.2.0` (`CVE-2026-116684` / `GHSA-w4fh-qvv9-3v23`, `CVE-2026-116685` / `GHSA-5h93-6whr-6q8j`, `CVE-2026-116686` / `GHSA-rwrf-2pqf-9j8j`)
  • PyPI `strands-agents < 1.48.0` (`GHSA-3w55-933p-8pv5` symlink LPE and `GHSA-763v-5pjm-fx36` world-readable `/tmp/strands/sessions` storage) and `strands-agents < 1.43.0` / npm `@strands-agents/sdk < 1.5.0` running multi-caller Agent2Agent (`A2AServer`) endpoints (`CVE-2026-115419` / `GHSA-4v6f-7vjh-6x4m`)
  • PyPI `awslabs.aws-api-mcp-server >= 1.3.0, < 1.6.0` using `streamable-http` OAuth authentication (`CVE-2026-115413` / `GHSA-4prq-p2x5-w9fh`), and PyPI `google-adk >= 1.19.0, < 1.39.1` and `>= 2.0.0a1, < 2.7.0` sharing a `FileArtifactService` root across multiple apps (`CVE-2026-115522`)
  • PyPI `langgraph-api >= 0.1.2, < 0.15.0` (`CVE-2026-118668` / `GHSA-4hm6-w6qq-w73v`), `langgraph-sdk >= 0.4.0, < 0.4.6` / `langgraph >= 1.2.3, < 1.2.14` (`CVE-2026-118667`), `langgraph-cli >= 0.4.15, < 0.4.32` (`CVE-2026-115885`), `langgraph-checkpoint-postgres < 3.1.3` / `langgraph-checkpoint-sqlite < 3.1.2` (`CVE-2026-115883`), and PyPI `mem0ai < 2.0.18` / npm `mem0ai < 3.1.6` (`CVE-2026-116575` / `GHSA-5324-3w58-qpp2`)
  • PyPI `fastmcp` and `fastmcp-slim` `>= 3.0.0, < 3.4.10` and `>= 4.0.0, < 4.1.2` using component `auth=` checks with transforms (`GHSA-7xgj-8392-5q4r`) or `ClientCredentialsOAuthProvider` / `PrivateKeyJWTOAuthProvider` (`GHSA-vqqp-x5qm-5qp4`)

Update, 5 November 2026: On 4 November (00:17 UTC), Trend Micro’s Zero Day Initiative (ZDI) and NVD published three additional advisories for MLflow: CVE-2026-118623 (ZDI-26-839, CVSS 7.8 High, untrusted pickle deserialization in PickleEvaluationArtifact._save / _load when evaluating untrusted artifacts), CVE-2026-118627 (ZDI-26-843, CVSS 7.8 High, untrusted pickle deserialization in mlflow.shap._load_pyfunc / load_explainer when loading SHAP explainer models), and CVE-2026-118628 (ZDI-26-844, CVSS 8.2 High, unauthenticated server-side request forgery in the /get-artifacts handler via unvalidated artifact URIs).

Update, 29 October 2026: Late on 28 October (21:05 to 21:17 UTC), four additional advisories were published for LangGraph and Mem0 (mem0ai):

  • LangGraph API JavaScript custom-auth Store namespace bypass (CVE-2026-118668 / GHSA-4hm6-w6qq-w73v, CVSS 8.6 High, fixed in langgraph-api 0.15.0): In langgraph-api >= 0.1.2, < 0.15.0, when a deployment used JavaScript or TypeScript custom authentication and its Store authorization handler restricted tenants by rewriting the requested namespace rather than rejecting the request outright, the server failed to carry the rewritten namespace across the JS/Python runtime boundary before executing the Store operation. Any authenticated user could read, search, overwrite, or delete another user’s Store data.
  • LangGraph SDK v3 streaming path traversal (CVE-2026-118667 / GHSA-3fx2-cqw9-xr3c, Medium, fixed in langgraph-sdk 0.4.6 and langgraph 1.2.14) and langgraph deploy --verbose push-token leak (CVE-2026-115885 / GHSA-8prr-f582-jqq3, Medium, fixed in langgraph-cli 0.4.32): In langgraph-sdk 0.4.0 through 0.4.5, the v3 streaming client (get_tree and _fetch_state) interpolated thread_id and assistant_id into URL paths without calling _quote_path_param (commit d89c7e100e), letting relative path segments redirect authenticated GET requests to /store/items?.... Meanwhile, in langgraph-cli 0.4.15 through 0.4.31, running langgraph deploy --verbose echoed the docker login --password-stdin stdin stream, printing the temporary registry push token into CI logs (commit 1afaca35a0).
  • Mem0 _build_session_scope delimiter collision and cross-tenant conversation leak (CVE-2026-116575 / GHSA-5324-3w58-qpp2, CVSS 6.3 Medium, fixed in PyPI mem0ai 2.0.18 and npm mem0ai 3.1.6): Both mem0/memory/main.py (_build_session_scope) and mem0-ts/src/oss/src/memory/index.ts (buildSessionScope) joined agent_id, run_id, and user_id into "&".join(f"{key}={val}" ...) without forbidding or escaping & and =. A caller who controlled only run_id could pass run_id="proj-x&user_id=alice" to collide byte-for-byte with {"user_id": "alice", "run_id": "proj-x"}, causing Memory.add() to read Alice’s last 10 raw messages into the attacker’s LLM fact-extraction prompt and overwrite or evict Alice’s conversation history (commit 71f2ebefa3).

Update, 28 October 2026: Between 27 and 28 October, ten additional advisories were published for MLflow 3.17.0, the Model Context Protocol (MCP) Python SDK (mcp 1.30.0 and 2.2.0), Amazon Strands Agents SDK 1.48.0, AWS API MCP Server 1.6.0, and Google Agent Development Kit (google-adk 1.39.1 and 2.7.0):

  • Strands Agents FileSessionManager /tmp symlink LPE and world-readable session storage (GHSA-3w55-933p-8pv5, CVSS 7.1 High, and GHSA-763v-5pjm-fx36, Medium, fixed in strands-agents 1.48.0): In strands-agents < 1.48.0, FileSessionManager defaulted storage_dir to os.path.join(tempfile.gettempdir(), "strands/sessions") (/tmp/strands/sessions on Linux and macOS), created directories and JSON files under the default 022 umask (0755 / 0644), and wrote updates via open(f"{path}.tmp", "w") followed by os.replace() without O_NOFOLLOW. Any local user on a shared developer workstation, notebook server, or CI runner could read persisted prompts, tool outputs, and secrets (GHSA-763v-5pjm-fx36), pre-create a symlink at session.json.tmp to overwrite arbitrary files owned by the victim user, or tamper with session files for persistent prompt injection (GHSA-3w55-933p-8pv5, commit 8658b9e0d5). Version 1.48.0 moves the default directory to ~/.strands/sessions/ with mode 0700.
  • AWS API MCP Server OAuth token audience bypass (CVE-2026-115413 / GHSA-4prq-p2x5-w9fh, CVSS 7.8 High, fixed in awslabs.aws-api-mcp-server 1.6.0): When awslabs.aws-api-mcp-server (>= 1.3.0, < 1.6.0) ran with AWS_API_MCP_TRANSPORT="streamable-http" and AUTH_TYPE="oauth", it verified the bearer token’s JWKS signature, iss, and exp, but never checked aud or client_id / azp. Any token issued by the same identity provider (such as another app client in the same Amazon Cognito user pool) was accepted and allowed the caller to execute AWS CLI commands under the server’s host IAM credentials. Version 1.6.0 (commit b97eefc7c2) requires setting either AUTH_AUDIENCE or AUTH_CLIENT_ID alongside AUTH_ISSUER and AUTH_JWKS_URI.
  • Google Agent Development Kit (google-adk) cross-app FileArtifactService leak (CVE-2026-115522, CVSS v4.0 5.3 Medium, fixed in 1.39.1 and 2.7.0): In google-adk 1.19.0 before 1.39.1 and 2.0.0a1 before 2.7.0, FileArtifactService stored artifacts under root/users/{user_id}/... and dropped app_name from the path entirely, allowing one application to list, read, overwrite, or delete artifacts belonging to another application sharing the same artifact root (commit f72f0db58c).
  • MLflow unauthenticated third-party scorer artifact RCE (CVE-2026-116692 / GHSA-26p8-2jq9-3vq9, CVSS 9.8 Critical): In mlflow 3.12.0rc0 through 3.16.x, Scorer.model_validate() checked module_path == m or module_path.startswith(m + ".") against THIRD_PARTY_SCORER_ALLOWED_MODULES (mlflow.genai.scorers.trulens, ragas, deepeval, phoenix) and then ran importlib.import_module(module_path). By creating a file-backed experiment with artifact_location: "file:///usr/local/lib/python3.10/site-packages/mlflow/genai/scorers/trulens", uploading a .py file as a run artifact, and calling POST /ajax-api/3.0/mlflow/scorer/invoke with module: "mlflow.genai.scorers.trulens.<run_id>.artifacts.<name>", an unauthenticated caller executed arbitrary Python code as root on the official ghcr.io/mlflow/mlflow container image (commit a732443e69).
  • MLflow prompt model-version arbitrary file read (CVE-2026-116690 / GHSA-m65p-9j7q-6w33, CVSS 7.5 High) and CrewAI autolog API-key leakage (CVE-2026-116691 / GHSA-j7qq-vr8c-xvw3, CVSS 6.5 Medium): In mlflow 3.10.0 through 3.16.1, the CreateModelVersion prompt branch (mlflow.prompt.is_prompt tag) rejected file:// URIs and leading / paths but skipped _validate_non_local_source_contains_relative_paths() whenever parsed.scheme was empty, allowing source: "../../../etc" or "." to serve arbitrary server files (including basic_auth.db) via /api/2.0/mlflow/model-versions/get-artifact (commit 81730a2ce4). Separately, in mlflow 2.19.0 through 3.16.x, mlflow.crewai.autolog() serialized Agent.__dict__ into span attributes and mlflow.spanInputs, exporting plaintext api_key='sk-proj-...' strings from CrewAI LLM objects across 5 span locations (commit 20927930d2).
  • MCP Python SDK (mcp 1.30.0 / 2.2.0):
    • CVE-2026-116684 (GHSA-w4fh-qvv9-3v23, CVSS 6.8 Medium): BearerAuthBackend never checked AccessToken.resource (RFC 8707 audience) against AuthSettings.resource_server_url. Upgrading alone does not enable the fix: you must set validate_token_resource=True in AuthSettings and return the token’s audience in AccessToken.resource from verify_token() (commit 0c9136841f).
    • CVE-2026-116686 (GHSA-rwrf-2pqf-9j8j, CVSS 6.5 Medium): When validating call_tool results against a server’s outputSchema, ClientSession._validate_tool_result() called jsonschema.validate() without an isolated referencing.Registry(), allowing a malicious MCP server to place http://, https://, or file:// $ref URIs in outputSchema to trigger synchronous SSRF, local file probing, or event-loop hangs on the client (commit 5ebdfed406).
    • CVE-2026-116685 (GHSA-5h93-6whr-6q8j, CVSS 5.9 Medium): create_mcp_http_client() defaulted to follow_redirects=True, causing Streamable HTTP, SSE, and OAuth client flows to forward custom credential headers (X-API-Key) and 307/308 request bodies across cross-origin redirects (commit 2106335745).

TL;DR

Between 22 and 23 October 2026, four major AI agent and evaluation frameworks (Amazon Strands Agents SDK, LangGraph, Prefect FastMCP, and MLflow) disclosed five vulnerabilities affecting multi-tenant session isolation, MCP tool authorization, OAuth client-credential binding, and tracking-server egress:

  1. Amazon Strands Agents SDK (CVE-2026-115419 / GHSA-4v6f-7vjh-6x4m, CVSS 8.1 High): A2AServer and StrandsA2AExecutor accepted a single Agent instance and reused its in-memory conversation state and streaming state across every caller’s Agent2Agent (A2A) context_id. Any caller could read another tenant’s conversation history or inject persistent prompt instructions into subsequent sessions.
  2. LangGraph Checkpoint Postgres & SQLite (CVE-2026-115883 / GHSA-6hmw-9432-67qq, CVSS 6.0 Medium): PostgresStore and SqliteStore serialize a namespace tuple ("tenant", "bob") by joining labels with a period (".".join(namespace) -> "tenant.bob"). BaseStore only rejected . inside labels on put() / aput(), leaving get(), search(), delete(), list_namespaces(), and batch() / abatch() (including PutOp items inside batch()) unchecked. Passing ("tenant.bob",) read, overwrote, or deleted ("tenant", "bob").
  3. Prefect FastMCP (GHSA-7xgj-8392-5q4r, CVSS 6.8 Medium, and GHSA-vqqp-x5qm-5qp4, CVSS 5.9 Medium): First, when a ToolTransform or custom Transform added tags or modified metadata that an AuthCheck (such as restrict_tag("admin", scopes=["admin"])) relied on, list_tools() hid the tool, but direct call_tool() / get_resource() lookups checked the pre-transform component or returned the transformed component without running _component_auth_allows(). Second, ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider never exposed or forwarded the issuer= parameter to the underlying MCP SDK, so a compromised or user-supplied MCP server could advertise its own OAuth token endpoint and capture the client’s client_id and client_secret.
  4. MLflow (CVE-2026-115886 / GHSA-2wqp-q3c6-ccj5, CVSS 6.5 Medium): The unauthenticated POST /ajax-api/3.0/mlflow/scorer/invoke endpoint deserialized third_party_scorer_data via Scorer.model_validate_json(), reflectively instantiating mlflow.genai.scorers.trulens classes and forwarding raw kwargs (api_base, base_url, api_key, completion_kwargs) straight into LiteLLM(), allowing unauthenticated SSRF with caller-controlled headers.

What to do now:

  • Upgrade strands-agents to >= 1.43.0 (@strands-agents/sdk to >= 1.5.0) and migrate A2AServer(agent=...) calls to A2AServer(agent_factory=lambda context_id: ...).
  • Upgrade langgraph-checkpoint-postgres to >= 3.1.3, langgraph-checkpoint-sqlite to >= 3.1.2, and langgraph-checkpoint to >= 4.3.0.
  • Upgrade fastmcp / fastmcp-slim to >= 4.1.2 (or >= 3.4.10 on 3.x), and explicitly pass issuer="https://..." to ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider.
  • Upgrade mlflow to >= 3.17.0.

Affected packages and projects

EcosystemPackageCVE / GHSASeverityAffected versionsFixed versionsAction required beyond upgrade?
PyPI / npmstrands-agents, @strands-agents/sdkCVE-2026-115419 (GHSA-4v6f-7vjh-6x4m)CVSS 8.1 Highstrands-agents < 1.43.0; @strands-agents/sdk < 1.5.01.43.0 (PyPI), 1.5.0 (npm)Yes (recommended): pass agent_factory= (agentFactory) to A2AServer so each context_id gets its own Agent and session_manager
PyPIstrands-agentsGHSA-3w55-933p-8pv5, GHSA-763v-5pjm-fx36CVSS 7.1 High / Medium< 1.48.01.48.0 (1.62.0 current)Remove legacy /tmp/strands/sessions directories on shared hosts
PyPIawslabs.aws-api-mcp-serverCVE-2026-115413 (GHSA-4prq-p2x5-w9fh)CVSS 7.8 High>= 1.3.0, < 1.6.01.6.0Yes: configure AUTH_AUDIENCE or AUTH_CLIENT_ID when AUTH_TYPE=oauth
PyPIgoogle-adkCVE-2026-115522CVSS v4.0 5.3 Medium>= 1.19.0, < 1.39.1; >= 2.0.0a1, < 2.7.01.39.1, 2.7.0 (2.13.0 current)Pre-upgrade artifacts under root/users/ remain readable until migrated or deleted
PyPIlanggraph-checkpoint-postgres, langgraph-checkpoint-sqlite, langgraph-checkpointCVE-2026-115883 (GHSA-6hmw-9432-67qq)CVSS 6.0 Mediumlanggraph-checkpoint-postgres < 3.1.3; langgraph-checkpoint-sqlite < 3.1.2; langgraph-checkpoint < 4.3.03.1.3, 3.1.2, 4.3.0Custom external stores that join namespaces with . should call validate_op_namespace(op)
PyPIfastmcp, fastmcp-slimGHSA-7xgj-8392-5q4rCVSS 6.8 Medium>= 3.0.0, < 3.4.10; >= 4.0.0, < 4.1.23.4.10, 4.1.2No
PyPIfastmcp, fastmcp-slimGHSA-vqqp-x5qm-5qp4CVSS 5.9 Medium>= 4.0.0a1, < 4.1.24.1.2 (requires mcp >= 2.2.0)Yes: pass issuer="https://..." to ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider, and rotate secrets exposed to untrusted MCP URLs
PyPImcpCVE-2026-116684 (GHSA-w4fh-qvv9-3v23)CVSS 6.8 Medium>= 1.10.0, < 1.30.0; >= 2.0.0, < 2.2.01.30.0, 2.2.0Yes: set validate_token_resource=True in AuthSettings and populate AccessToken.resource in verify_token()
PyPImcpCVE-2026-116686 (GHSA-rwrf-2pqf-9j8j), CVE-2026-116685 (GHSA-5h93-6whr-6q8j)CVSS 6.5 / 5.9 Medium>= 1.8.0, < 1.30.0; >= 2.0.0, < 2.2.01.30.0, 2.2.0No (cross-origin redirects now fail unless endpoint URL is updated to final target)
PyPImlflowCVE-2026-116692 (GHSA-26p8-2jq9-3vq9)CVSS 9.8 Critical>= 3.12.0rc0, < 3.17.03.17.0 (3.17.2 current)Audit writable site-packages/mlflow/genai/scorers/* directories for rogue run artifact subtrees
PyPImlflowCVE-2026-116690 (GHSA-m65p-9j7q-6w33)CVSS 7.5 High>= 3.10.0, < 3.17.03.17.0Audit registered model versions with mlflow.prompt.is_prompt tags for relative source paths
PyPImlflowCVE-2026-116691 (GHSA-j7qq-vr8c-xvw3)CVSS 6.5 Medium>= 2.19.0, < 3.17.03.17.0Yes: rotate LLM provider keys used with mlflow.crewai.autolog() and purge recorded spans
PyPImlflowCVE-2026-115886 (GHSA-2wqp-q3c6-ccj5)CVSS 6.5 Medium>= 3.12.0rc0, < 3.17.03.17.0No (previously stored scorer rows with destination kwargs are also blocked on read)

Technical analysis

1. Strands Agents SDK (CVE-2026-115419): shared Agent state across A2A context_ids

In Amazon’s Strands Agents SDK (strands-agents on PyPI and @strands-agents/sdk on npm), A2AServer adapts a Strands Agent to the Agent2Agent (A2A) protocol. Before Python 1.43.0 (commit 398343faad) and TypeScript 1.5.0 (commit f8ba9376fe), A2AServer accepted a single agent: SAAgent instance and passed that same object into StrandsA2AExecutor.

Because a Strands Agent holds its conversation messages (agent.messages), system prompt state, and session manager on the instance itself (and StrandsA2AExecutor also stored _current_artifact_id and _is_first_chunk on self), every incoming A2A request mutated the exact same conversation history regardless of the request’s context_id:

  • Cross-tenant conversation disclosure: Caller B sending a request with a fresh context_id caused the model to run with Caller A’s prior turns still in agent.messages.
  • Cross-session prompt injection: Any caller could plant persistent instructions in the shared agent.messages list that steered tool calls and responses for subsequent callers.

The fix introduces agent_factory: Callable[[str], Agent] (agentFactory: (contextId: string) => InvokableAgent in TypeScript) alongside an LRU context map (max_contexts = 1000) where each context_id owns a dedicated _ContextEntry(agent, lock) and per-request _StreamState:

# strands-py/src/strands/multiagent/a2a/server.py (patched in 1.43.0)
from strands import Agent
from strands.multiagent.a2a import A2AServer

# DeprecationWarning if a single `agent=` is passed; use agent_factory for multi-caller isolation:
server = A2AServer(
    agent_factory=lambda context_id: Agent(
        model=my_model,
        tools=[...],
        # Wire any session_manager per context_id here
    ),
    max_contexts=1000,
)

If legacy code still passes a single agent= instance, 1.43.0 / 1.5.0 emits a DeprecationWarning, serializes all requests behind a single lock, and swaps take_snapshot() / load_snapshot() state onto and off the shared instance per context_id (provided no shared session_manager is attached). For real concurrency and multi-tenant isolation, switch to agent_factory.

2. LangGraph (CVE-2026-115883): dot-joined namespace collision in PostgresStore and SqliteStore

In LangGraph’s persistence layer, namespaces are passed as tuples of strings such as ("org_123", "user_456"). Both PostgresStore (langgraph-checkpoint-postgres) and SqliteStore (langgraph-checkpoint-sqlite) store that tuple in SQL as a single dot-joined prefix column (".".join(namespace) -> "org_123.user_456").

To prevent ("org_123.user_456",) from colliding with ("org_123", "user_456"), BaseStore._validate_namespace() forbade periods (.) in namespace labels. However, before commit bfcfea554e, BaseStore only called _validate_namespace() inside put() and aput():

  • store.get(namespace, key), store.search(namespace_prefix), store.delete(namespace, key), and store.list_namespaces(prefix=..., suffix=...) (plus aget, asearch, adelete, alist_namespaces) never validated the labels.
  • store.batch(ops) and store.abatch(ops) validated nothing at all, not even PutOp(namespace, key, value) items passed directly inside batch().

Consequently, if an application built a namespace tuple from a caller-supplied string tenant_id (for example namespace = (f"tenant_{user_input}",) or (user_input,)), passing "tenant_a.secret_ns" allowed the caller to:

# Against langgraph-checkpoint-postgres < 3.1.3 or langgraph-checkpoint-sqlite < 3.1.2:
store.get(("tenant_a.secret_ns",), "api_keys")          # reads ("tenant_a", "secret_ns")
store.delete(("tenant_a.secret_ns",), "api_keys")       # deletes ("tenant_a", "secret_ns")
store.batch([PutOp(("tenant_a.secret_ns",), "k", {})])  # overwrites ("tenant_a", "secret_ns")

Commit bfcfea554e adds _validate_namespace_labels() to get, search, delete, list_namespaces, and their async counterparts in langgraph-checkpoint 4.3.0 (before queueing into AsyncBatchedBaseStore._aqueue, so one caller’s invalid label cannot fail concurrent callers sharing the batch), and calls validate_op_namespace(op) inside _group_ops() in both PostgresStore (3.1.3) and SqliteStore (3.1.2):

--- a/libs/checkpoint-postgres/langgraph/store/postgres/base.py
+++ b/libs/checkpoint-postgres/langgraph/store/postgres/base.py
@@ -1386,6 +1387,7 @@ def _group_ops(ops: Iterable[Op]) -> tuple[dict[type, list[tuple[int, Op]]], int
     grouped_ops: dict[type, list[tuple[int, Op]]] = defaultdict(list)
     tot = 0
     for idx, op in enumerate(ops):
+        validate_op_namespace(op)
         grouped_ops[type(op)].append((idx, op))
         tot += 1
     return grouped_ops, tot

3. FastMCP (GHSA-7xgj-8392-5q4r and GHSA-vqqp-x5qm-5qp4): post-transform auth bypass and un-bound M2M OAuth client secrets

Prefect’s fastmcp 4.1.2 and 3.4.10 (commit 85df42fbf2) fix two distinct security flaws:

A. Direct component calls skipped post-transform AuthCheck (GHSA-7xgj-8392-5q4r)

FastMCP allows servers to attach component-level authorization rules (such as auth=restrict_tag("admin", scopes=["admin"])) and also attach transforms (ToolTransform, Transform) that rename components or add tags like tags={"admin"} at the provider, server, or session layer.

When a client called list_tools(), FastMCP ran all transforms first and then filtered the transformed list through _tool_auth_allows(tool), hiding the "admin"-tagged tool from unprivileged callers. However, on the direct lookup path (server.call_tool("tagged_tool"), get_resource(), get_prompt()), FastMCP checked authorization on the untransformed component or returned the transformed component from call_next without re-evaluating _component_auth_allows() after the transform pipeline finished. Furthermore, if a concrete resource (secret://data) was denied by its AuthCheck and a wildcard resource template (secret://{name}) existed on the same server or via a ProxyProvider, falling through on denial allowed the template to serve the protected URI anyway.

The fix in 85df42fbf2 re-runs _component_auth_allows() on the final transformed Tool, Resource, ResourceTemplate, and Prompt at every lookup layer, and ensures a denied concrete resource shadows any matching resource template at that URI.

B. M2M OAuth wrappers omitted issuer binding (GHSA-vqqp-x5qm-5qp4)

In FastMCP 4.0.0a1 through 4.1.1, ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider (fastmcp/client/auth/client_credentials.py) wrapped the upstream mcp Python SDK’s machine-to-machine OAuth providers, which discover the MCP server’s authorization server via RFC 9728 Protected Resource Metadata.

Although the underlying mcp SDK supported an issuer= parameter to verify that the authorization server advertised by the MCP server matched the issuer that actually owns the client’s client_id and client_secret, FastMCP’s wrappers omitted issuer from __init__() and _bind(). If a FastMCP client connected to a user-supplied or compromised MCP server URL, that server could advertise https://attacker[.]example/oauth/token in its metadata and receive the client’s client_id and client_secret in the token-exchange POST.

FastMCP 4.1.2 bumps its dependency floor to mcp>=2.2.0, adds issuer: str | None = None to both providers, isolates cached tokens by (client_id, scopes, issuer), and emits an MCPDeprecationWarning when issuer is omitted:

from fastmcp import Client
from fastmcp.client.auth import ClientCredentialsOAuthProvider

# Upgrading to fastmcp >= 4.1.2 is not enough by itself; pass issuer= explicitly:
auth = ClientCredentialsOAuthProvider(
    client_id="my-client-id",
    client_secret="my-client-secret",
    scopes=["read", "write"],
    issuer="https://auth.example.com",
)

4. MLflow (CVE-2026-115886): unauthenticated POST /ajax-api/3.0/mlflow/scorer/invoke SSRF via third_party_scorer_data.kwargs

In MLflow 3.12.0rc0 through 3.16.x, the GenAI evaluation routes POST /ajax-api/3.0/mlflow/scorer/invoke, POST /ajax-api/3.0/mlflow/scorers/register, and POST /ajax-api/3.0/mlflow/genai/evaluate accept a JSON serialized_scorer string. Notably, /ajax-api/3.0/mlflow/scorer/invoke is registered without an authentication decorator on default MLflow tracking servers and runs Scorer.model_validate_json(serialized_scorer) synchronously.

Inside mlflow/genai/scorers/base.py, Scorer.model_validate_json() checked module_path against THIRD_PARTY_SCORER_ALLOWED_MODULES (ragas, deepeval, trulens, phoenix), imported the module, looked up getattr(module, class_name), and called scorer_class(**init_kwargs) with unconstrained data["kwargs"]. When trulens and trulens-providers-litellm were installed in the server environment, selecting module: "mlflow.genai.scorers.trulens" with a non-native model URI such as "openai:/gpt-4o" routed **kwargs directly into trulens.providers.litellm.LiteLLM(model_engine=..., **kwargs), which passed api_base, base_url, api_key, or completion_kwargs into litellm.completion().

Commit bac30b4148 (released in MLflow 3.17.0) adds _validate_serialized_scorer_payload() and _validate_third_party_scorer_data() in mlflow/utils/validation.py, recursively walking every nested third_party_scorer_data dictionary and list of 2-item pairs (used by completion_kwargs) and rejecting any key in GATEWAY_DESTINATION_KEYS (api_base, base_url, api_key, custom_llm_provider, litellm_params, etc.) as well as call_source decorator scorers:

--- a/mlflow/utils/validation.py
+++ b/mlflow/utils/validation.py
@@ -1119,6 +1119,25 @@
+def _find_destination_keys(value: Any) -> set[str]:
+    if isinstance(value, dict):
+        found = GATEWAY_DESTINATION_KEYS & set(value)
+        for nested in value.values():
+            found |= _find_destination_keys(nested)
+        return found
+    if isinstance(value, list):
+        found: set[str] = set()
+        for nested in value:
+            if isinstance(nested, list) and len(nested) == 2 and isinstance(nested[0], str):
+                found |= GATEWAY_DESTINATION_KEYS & {nested[0]}
+            found |= _find_destination_keys(nested)
+        return found
+    return set()

5. MLflow (CVE-2026-116692, CVE-2026-116690, and CVE-2026-116691): third-party scorer artifact RCE, prompt source file read, and CrewAI autolog key leak

Three further MLflow 3.17.0 advisories were published on 27 October:

  1. Unauthenticated RCE via third-party scorer prefix matching (CVE-2026-116692 / GHSA-26p8-2jq9-3vq9, CVSS 9.8 Critical): In mlflow/genai/scorers/base.py, Scorer.model_validate() verified module_path with a prefix match:
    if not any(
        module_path == m or module_path.startswith(m + ".")
        for m in THIRD_PARTY_SCORER_ALLOWED_MODULES
    ):
    Because mlflow server allows creating an experiment with a custom artifact_location (such as file:///usr/local/lib/python3.10/site-packages/mlflow/genai/scorers/trulens, which is writable by the default uid=0 process in ghcr.io/mlflow/mlflow:v3.14.0), an unauthenticated attacker could create a run <run_id>, upload payload.py via the artifact API, and call POST /ajax-api/3.0/mlflow/scorer/invoke with module: "mlflow.genai.scorers.trulens.<run_id>.artifacts.payload". Scorer.model_validate_json() imported the uploaded module and executed its class constructor before checking trace IDs. Commit a732443e69 replaces the prefix check with an exact membership check (if module_path not in THIRD_PARTY_SCORER_ALLOWED_MODULES:) and validates class_name against a closed per-module allowlist before calling importlib.import_module().
  2. Arbitrary file read via schemeless prompt model-version source (CVE-2026-116690 / GHSA-m65p-9j7q-6w33, CVSS 7.5 High): When creating a model version carrying the mlflow.prompt.is_prompt tag, _create_model_version() in mlflow/server/handlers.py rejected parsed.scheme == "file" and source.startswith("/"), but only called _validate_non_local_source_contains_relative_paths(source) when parsed.scheme was non-empty. Passing a schemeless relative path (source: "../../../etc" or ".") stored the traversal path directly as the model version’s artifact root, and GET /api/2.0/mlflow/model-versions/get-artifact?name=pwned-model&version=1&path=passwd served /etc/passwd or basic_auth.db. Commit 81730a2ce4 restricts schemeless prompt sources to _PROMPT_SOURCE_PLACEHOLDERS ("dummy-source" and "prompt-template").
  3. CrewAI autologging API-key disclosure (CVE-2026-116691 / GHSA-j7qq-vr8c-xvw3, CVSS 6.5 Medium): In mlflow/crewai/autolog.py, _get_agent_attributes() and _construct_full_inputs() expanded Agent.__dict__ and converted agent.llm, agent.function_calling_llm, and agent.agent_executor to strings. Because CrewAI’s LLM.__repr__() includes api_key='sk-proj-...', enabling mlflow.crewai.autolog() automatically exported plaintext LLM provider API keys into 5 span attributes (attributes.llm, attributes.function_calling_llm, attributes.agent_executor, and two attributes.mlflow.spanInputs entries). Commit 20927930d2 reduces LLM instances to their model name and redacts keys matching _SENSITIVE_KEY_PATTERN.

6. Model Context Protocol Python SDK (mcp 1.30.0 / 2.2.0): CVE-2026-116684, CVE-2026-116686, and CVE-2026-116685

In modelcontextprotocol/python-sdk:

  • Opt-in RFC 8707 audience check (CVE-2026-116684 / GHSA-w4fh-qvv9-3v23, commit 0c9136841f): BearerAuthBackend in src/mcp/server/auth/middleware/bearer_auth.py accepted any AccessToken returned by token_verifier.verify_token(token) without comparing AccessToken.resource against AuthSettings.resource_server_url. In mcp 1.30.0 and 2.2.0, you must explicitly opt in by setting validate_token_resource=True on AuthSettings and populating AccessToken.resource inside verify_token().
  • Tool outputSchema $ref SSRF and file read (CVE-2026-116686 / GHSA-rwrf-2pqf-9j8j, commit 5ebdfed406): In src/mcp/client/session.py, _validate_tool_result() called jsonschema.validate(result.structuredContent, output_schema) with jsonschema’s default resolver, which fetches unresolved $ref URLs via urllib.request.urlopen() (http://, https://, and file://) synchronously on the event loop without a timeout. 1.30.0 and 2.2.0 pass an explicit empty registry=Registry() from referencing so only in-document #/$defs/... references resolve.
  • Same-origin redirect restriction (CVE-2026-116685 / GHSA-5h93-6whr-6q8j, commit 2106335745): create_mcp_http_client() no longer enables follow_redirects=True; request_within_origin() and stream_within_origin() follow redirects only when the target stays on the same origin (or upgrades http -> https on the same host).

Detection and remediation

1. Audit environments, lockfiles, and vulnerable call patterns

Run the following commands across your Python and TypeScript repositories:

# 1. Check installed Python package versions
pip show strands-agents langgraph-checkpoint langgraph-checkpoint-postgres langgraph-checkpoint-sqlite fastmcp fastmcp-slim mlflow 2>/dev/null || true

# 2. Check npm lockfiles for @strands-agents/sdk < 1.5.0
npm ls @strands-agents/sdk 2>/dev/null || true

# 3. Audit code for single-agent A2AServer(...) construction and FastMCP M2M OAuth without issuer=
rg -n 'A2AServer\s*\(|StrandsA2AExecutor\s*\(|A2AExecutor\s*\(' --glob '*.{py,ts,js}' . 2>/dev/null
rg -n '(ClientCredentialsOAuthProvider|PrivateKeyJWTOAuthProvider)\s*\(' --glob '*.py' . 2>/dev/null

2. Upgrade packages and apply required code changes

  1. Strands Agents SDK (CVE-2026-115419):
    • Upgrade Python strands-agents to >= 1.43.0 and TypeScript @strands-agents/sdk to >= 1.5.0.
    • Refactor every multi-caller A2AServer(agent) / StrandsA2AExecutor(agent) call to pass agent_factory (agentFactory in TypeScript) so each context_id receives a fresh Agent instance and a context-scoped session_manager.
  2. LangGraph (CVE-2026-115883):
    • Upgrade langgraph-checkpoint-postgres to >= 3.1.3, langgraph-checkpoint-sqlite to >= 3.1.2, and langgraph-checkpoint to >= 4.3.0 (or upgrade self-hosted LangSmith Agent Server to >= 0.15.5 / >= 0.16.2).
  3. FastMCP (GHSA-7xgj-8392-5q4r and GHSA-vqqp-x5qm-5qp4):
    • Upgrade fastmcp / fastmcp-slim to >= 4.1.2 (or >= 3.4.10 on the 3.x branch).
    • Add issuer="https://<your-expected-auth-server>" to every ClientCredentialsOAuthProvider and PrivateKeyJWTOAuthProvider instantiation, and rotate any client_secret that was used with untrusted or user-configured MCP server URLs.
  4. MLflow (CVE-2026-115886):
    • Upgrade mlflow to >= 3.17.0, and audit tracking-store registered_scorers rows and HTTP logs for requests to /ajax-api/3.0/mlflow/scorer/invoke containing third_party_scorer_data.

Timeline

  • 2026-10-22 (20:04 UTC): AWS publishes bulletin 2026-146-AWS and CVE-2026-115419 (GHSA-4v6f-7vjh-6x4m) alongside strands-agents 1.43.0 and @strands-agents/sdk 1.5.0.
  • 2026-10-23 (21:12 UTC): LangChain publishes CVE-2026-115883 (GHSA-6hmw-9432-67qq) and releases langgraph-checkpoint 4.3.0, langgraph-checkpoint-postgres 3.1.3, and langgraph-checkpoint-sqlite 3.1.2.
  • 2026-10-23 (21:17 UTC): MLflow publishes CVE-2026-115886 (GHSA-2wqp-q3c6-ccj5), fixed in mlflow 3.17.0.
  • 2026-10-23 (21:31 to 21:32 UTC): Prefect publishes GHSA-vqqp-x5qm-5qp4 and GHSA-7xgj-8392-5q4r for fastmcp and fastmcp-slim, fixed in 3.4.10 and 4.1.2.
  • 2026-10-27 (21:05 to 21:26 UTC): The GitHub Advisory Database and NVD publish CVE-2026-116684, CVE-2026-116685, and CVE-2026-116686 for the MCP Python SDK (mcp 1.30.0 and 2.2.0) alongside CVE-2026-116692 (CVSS 9.8), CVE-2026-116690, and CVE-2026-116691 for mlflow 3.17.0.

From research to remediation

Check whether this pattern exists in your codebase

Turn this research into a remediation workflow. Scan dependencies and package manifests for similar supply-chain risk, then prioritize fixes with reachability context.

References