NLTK's pathsec module includes an SSRF filter intended to block access to loopback, private, link-local, and multicast IP ranges. However, this filter can be bypassed via DNS rebinding, where the hostname resolves to a public IP during validation but resolves to an internal or loopback IP at connection time. This bypass defeats the strict ENFORCE mode, allowing unauthorized access to internal services. The vulnerability is identified as CVE-2026-12075 and has a high severity score of 8.6. A patch is available to address this issue.
NLTK versions prior to 3.9.3 contain an eval injection vulnerability in the nltk.collocations module. This flaw allows attackers controlling command-line arguments to execute arbitrary Python code by passing unsanitized input to eval() in the __main__ block of collocations.py. The vulnerability arises because command-line arguments are used as suffixes of BigramAssocMeasures without validation, enabling escape from intended attribute lookups and execution of OS commands via the os module. A patch is available to address this issue.
CVE-2026-73557 is a race condition vulnerability in the vllm-project's vllm software affecting versions from 0.20.2rc0 up to but not including 0.26.0. It arises from improper synchronization of a process-global sparse tensor invariant guard in PyTorch 2.11.0 during concurrent prompt-embedding reconstruction within a single chat completion request. This can lead to invalid sparse tensor deserialization and subsequent unsafe dense tensor conversion. The vulnerability requires the --enable-prompt-embeds feature (default off) but does not require multiple workers or multimodal embeddings. The flaw allows a malicious payload to bypass the follow-up guard, potentially causing memory corruption or denial of service as documented in the related CVE-2025-62164. The CVSS 4.0 base score is 6.3 (medium severity).
CVE-2026-73556 is a denial-of-service vulnerability in the vllm project's lm-format-enforcer backend. The backend compiles attacker-supplied regular expressions without a timeout or buildability check, allowing a crafted regex to stall the engine worker by consuming excessive CPU. This issue affects versions of vllm prior to 0.26.0 when the lm-format-enforcer backend is explicitly enabled. The default backend is not affected. The vulnerability results in unauthenticated denial of service by hanging the structured-output compile step and blocking concurrent requests.
vLLM versions prior to 0.26.0 have a vulnerability where malformed API requests cause error messages to leak sensitive internal information such as file paths, line numbers, OS username, home directory, Python version, and internal package structure. This occurs because the error handler converts validation exceptions to strings that include traceback details, which are insufficiently sanitized before being returned in HTTP responses. This information disclosure can aid attackers in fingerprinting the environment and crafting targeted exploits. The issue affects all POST endpoints accepting JSON bodies. A fix is available in vLLM 0.26.0 that changes the error handling to avoid leaking such details.
The atomic-agents-stack dashboard HTTP server has a path traversal vulnerability in its file serving component. It constructs filesystem paths directly from HTTP request paths without proper containment checks, allowing attackers to read arbitrary files outside the intended directory. By default, the server binds to loopback, but it can be configured to bind to all interfaces, exposing the vulnerability to the LAN. This flaw affects all versions prior to 1.1.0. A fix is available that enforces safe path resolution and restricts host binding.
MCP-for-Stata is a MCP server for integrating Stata into agent loops with a safety-first design. Prior to version 1.19.0, the ado_package_install MCP tool in stata-mcp concatenates user-controlled input directly into a Stata command string without any validation or sanitization. An attacker who can invoke the MCP tool or the equivalent Python API can embed newline characters in the package argument to inject arbitrary Stata commands. Because Stata supports a shell escape command, this leads to full OS-level arbitrary command execution (RCE) under the account running the Stata-MCP server. The tool is registered in the default all profile, so no non-default configuration is required. This issue has been patched in version 1.19.0.
Material for MkDocs is a powerful documentation framework built on top of MkDocs. From 7.2.0 until 9.7.7, the mountSearchSuggest function in src/templates/assets/javascripts/components/search/suggest/index.ts contains a DOM-based cross-site scripting vulnerability in the optional search.suggest feature that allows a crafted q URL parameter to execute JavaScript in a documentation site's origin after user interaction. This issue is fixed in version 9.7.7.
Apache Airflow's secrets masker did not mask `var.json` Variable values whose value is a dict in the Rendered Templates UI — the dict value failed an `isinstance(str)` guard — so a secret stored as a JSON Variable and referenced in a template via `var.json` was displayed in cleartext to any user with access to that task's Rendered Templates view. Users are advised to upgrade to apache-airflow 3.3.1 or later, which masks nested Variable values regardless of type.
Apache Airflow's serialization layer reconstructed exception nodes by calling `import_string()` on a class name taken from the serialized blob and instantiating it with arguments from the same blob, with no restriction on what could be imported. An operator's `executor_config` reaches that branch, so a Dag author could place a value there that causes an arbitrary callable to be imported and invoked -- for example `subprocess.check_output`, or `builtins.eval` on the `builtins`-prefixed variant. The code runs in the **Scheduler**, which reconstructs serialized Dags in its normal loop with no request involved, and in the **API server**, on any authenticated read of the Dag such as `GET /api/v2/dags/{dag_id}/details`. Both are components the Airflow security model states must never execute Dag-author code, and both hold the metadata database credentials and the JWT signing secret. No non-default configuration is required. This is a **different sink from CVE-2026-33264**, which covered only the trigger branch of the same deserializer: deployments that upgraded in response to that advisory are still affected through the exception branch and must upgrade again. Users are advised to upgrade to apache-airflow 3.3.1 or later, which restricts the imported class to a subclass of `BaseException`.
Apache Airflow's XCom `GET /api/v2/{...}/xcomEntries/{key}?deserialize=true` endpoint passed a string-literal payload through `BaseXCom.deserialize_value` without the `_check_forbidden_xcom_keys` guard, allowing an authenticated API user with XCom write-and-read access to instantiate arbitrary `airflow.*` classes on the API server (CWE-502). An authenticated user who can write an XCom value and then read it back with `deserialize=true` triggers the unsafe instantiation. Users are advised to upgrade to apache-airflow 3.3.1 or later, which rejects reserved XCom serialization keys submitted as JSON string literals.
Apache Airflow's secrets masker hides values stored under sensitive key names when they are displayed in the UI. The masker's recursion-depth limit did not descend into values nested inside a list, tuple, or set beyond that limit, so an Airflow Variable holding such a deeply-nested value was shown unmasked in the Variables UI. The exposure is limited to the UI: any authenticated user who can see the Variable in the UI can already read its full value through the Variables REST API, so this does not disclose data the user could not otherwise obtain — the masking is a shoulder-surfing defense for the UI, not an access-control boundary.
This is an incomplete-fix follow-up to CVE-2026-42358, whose fix made only the dictionary walk unbounded; lists, tuples, and sets beyond the depth limit remained unmasked in the UI. Deployments that applied the CVE-2026-42358 fix should also upgrade to address this residual case. Upgrade to apache-airflow 3.3.1 or later.
Apache Airflow 3.3.0 moved human-in-the-loop tasks from the triggerer to a new `awaiting_input` task state swept by the scheduler. That sweep deserializes the task instance's `next_kwargs` without an allow-list, so a Dag author — who controls that value through the task execution API — can cause an arbitrary module import and object instantiation inside the scheduler process, or terminate the scheduler job. No non-default configuration is required: the sweep runs unconditionally every 15 seconds, and the default `allowed_deserialization_classes` setting does not cover this code path. Versions before 3.3.0 are not affected, because human-in-the-loop tasks deferred onto the triggerer instead. This is a different code path from CVE-2026-58076, which covers the same unguarded exception-node deserialization reached elsewhere — deployments that applied that fix must upgrade for this issue as well. Users are advised to upgrade to apache-airflow 3.3.1 or later.
Apache Airflow's Task SDK rebuilt a `Callback` object from serialized data by re-running its constructor, which imports the module named by the stored callback path. Because `SyncCallback` is itself an Airflow class it passes the default `allowed_deserialization_classes` allow-list, so tightening that setting does not help. A Dag author — who controls a task instance's `next_kwargs` through the task execution API — can therefore cause an arbitrary module to be imported inside the scheduler process, when the scheduler's `awaiting_input` timeout sweep deserializes that value. No non-default configuration is required; the sweep runs unconditionally. Versions before 3.3.0 are not affected: the class existed, but the scheduler sweep that reaches it did not. This is a separate code path from CVE-2026-58076 and CVE-2026-67260, which cover different gadgets reaching deserialization — applying either of those fixes does not address this one. Users are advised to upgrade to apache-airflow 3.3.1 or later.
Apache Airflow's Config API did not mask team-scoped sensitive configuration values in multi-team deployments. When an administrator has enabled multi-team mode and exposed the Config API, an authenticated Viewer holding only configuration-read access — with no prior access to the secret — could read a team-scoped Celery broker URL, including its embedded credentials, in cleartext, while the equivalent global option was correctly masked. The secrets masker matched only base section and option names and did not normalize team-prefixed sections before the sensitivity check (CWE-200). This is a distinct masker bypass from CVE-2026-48828 and CVE-2026-48892: deployments that upgraded to apache-airflow 3.3.0 to address those issues remain affected by this team-scoped variant. Users are advised to upgrade to apache-airflow 3.3.1 or later, which normalizes team-scoped sections before masking.
Apache Airflow's Backfill API authorized a request against a Dag id supplied by the caller whenever the `backfill_id` path segment failed to parse. The authorization dependency parsed it with `int()` while the route handler parsed it as pydantic's `NonNegativeInt`, which accepts values `int()` rejects (`1.0` coerces to `1`); FastAPI resolves dependencies before endpoint validation, so the two acted on different Dags. An authenticated user holding edit permission on any single Dag could therefore read, pause and cancel backfills belonging to any other Dag, including moving another Dag's queued runs to `failed`. No non-default configuration is required and backfill ids are sequential, so finding a target is trivial. Users are advised to upgrade to apache-airflow 3.3.1 or later, which parses the backfill id with the same type the routes declare.
Apache Airflow wrote Variable values and Connection `extra` contents to the audit log in cleartext when they were submitted through the bulk endpoints (`PATCH /api/v2/variables` and `PATCH /api/v2/connections`). The audit-log masking recognised only top-level request fields, and a bulk request nests its entities two levels below, so no masking was applied to them. Any authenticated user with audit-log read access -- who need not hold Variables or Connections read at all -- could recover those secrets verbatim, and the Connection `extra` copy is stored unencrypted in the log while the connection table encrypts it. The Airflow UI's *Import Variables* action posts to this endpoint, so an ordinary operator import wrote every secret in the file to the log. This is a different code path from CVE-2026-50204: that fix shipped in 3.3.0 and covers the single-entity endpoints only, so deployments that upgraded in response to that advisory remain affected and must upgrade again. Users are advised to upgrade to apache-airflow 3.3.1 or later.
Apache Airflow's Task SDK did not mask the contents of a Variable whose JSON value is a list, so secrets stored in that shape appeared in cleartext in task logs and in the Rendered Templates UI. Masking was applied only when the deserialized value was a string or a dict; a list at the top level matched neither and was returned unmasked. Any authenticated user able to read the logs or rendered templates of a task that references such a Variable could recover the values, with no special configuration required. This is the list-shaped counterpart of CVE-2026-59244, whose fix covered the dict case only, so deployments that upgraded in response to that advisory remain affected and must upgrade again. Users are advised to upgrade to apache-airflow 3.3.1 or later.
Apache Airflow's asset materialization endpoint (`POST /api/v2/assets/{asset_id}/materialize`) and the XCom result check on `wait_dag_run_until_finished` authorized the target Dag without its team, unlike every other authorization site. A team-aware auth manager distinguishes a team-scoped Dag from a global one by that field -- the Keycloak auth manager, for example, checks the `DAG` resource instead of `DAG:<team>` -- so the team-scoped permission that should gate the request was never consulted. In a deployment running multi-team mode with a team-aware auth manager, an authenticated user in one team could trigger Dag runs belonging to another team, supplying their own `dag_run_id` and `conf`, and could read another team's XCom values. Deployments using the FAB auth manager are unaffected, as it has no multi-team support. Users are advised to upgrade to apache-airflow 3.3.1 or later, which resolves the Dag's team at both sites.
Apache Airflow's environment-variable secrets backend resolved a team-scoped Connection or Variable from the wrong team's scope. The guard meant to prevent this only ran when no team scope was supplied, and its pattern could not match a team name containing an underscore, which team names are allowed to contain. When the guard did not apply, the lookup fell through to an unconditional global read that resolved the stored `AIRFLOW_CONN__<TEAM>___<ID>` variable regardless of which team asked. In multi-team mode an authenticated user of one team could therefore have `POST /api/v2/connections/test` resolve another team's Connection and authenticate outward with that team's credentials; the endpoint uses the credentials rather than returning them. Exploitation requires `[core] multi_team` enabled, `[core] test_connection` set to `Enabled` (it ships `Disabled`), team-scoped secrets provisioned as environment variables in the API-server process, and knowledge of the encoded identifier. Redirecting the test at an attacker-controlled host is separately blocked. Users are advised to upgrade to apache-airflow 3.3.1 or later.