Code injection in Sony nnabla through 1.39.0 can execute attacker-supplied Python when a user runs the 'nnabla_cli encode_param' command on a malicious text-format parameter file. The affected releases are no longer supported, no vendor fix is available, and the disclosure references publicly available exploit code; exploitation depends on the operator choosing to process the crafted file, which keeps the assessed risk low.
Type confusion in Objenesis 3.5 and earlier can make newInstance() return an object of a different same-named class when one cached Objenesis/ObjenesisStd instance is shared across multiple class loaders. A local low-privileged attacker who can execute code in the same JVM as a co-tenant or plugin can win the name-keyed cache slot and cause object substitution across the trust boundary, or a repeatable ClassCastException crash when the caller uses the returned value as its own type. The risk is limited to multi-class-loader deployments that reuse a single cached instance; scores are low (CVSS 4.0 2.0; vendor CVSS 3.1 4.5), no CISA KEV entry or confirmed active exploitation is reported, and the advisory includes a proof-of-concept reproduction.
An out-of-bounds read in Qt Framework's QCborValue::fromCbor() can expose memory contents or crash a parsing application when malformed CBOR is deserialized. The assessed vector requires local access, high complexity, and user interaction, with exploitation depending on a Qt application parsing attacker-supplied CBOR; the exact trigger channel depends on the embedding application. No public exploit identified at time of analysis, and this is a low-priority issue, and the available signals agree on that; a vendor patch is available.
Memory corruption in libimobiledevice through 1.4.0 allows a local user with low-level privileges to corrupt memory in any process that passes a manipulated 'Value' argument into the afc_get_device_info_key function, with low impact on confidentiality, integrity and availability. The vendor was contacted early but did not respond, and no fixed release is identified in the available data, so exposure persists for anything built against 1.4.0 or earlier. A public write-up describing the flaw is referenced; it is not verified as working exploit code, and the aggregated decision signals (local-only path, low privileges required, low impact, no KEV listing) rate real-world risk as low and warrant monitoring rather than emergency action.
A race condition in HaschekSolutions PictShare through 3.7.1 allows an attacker who can drive the upload endpoints to swap a file between validation and storage, bypassing the mime-type, SHA-1 and naughty-list checks enforced in architect() and API::upload(). The flaw is reachable over the network but requires winning a narrow, timing-dependent TOCTOU window (AC:H, PR:L per the CVSS vector; vendor rates exploitability as difficult), and its practical impact is limited to integrity of stored uploads rather than code execution. A public write-up and PoC reference exist (xmit.xyz); it is not verified as working exploit code, there is no KEV listing, and the priority decision is monitor. Fixed in release 3.7.2 (commit 6306e33).
Command injection in apache-airflow-providers-ssh 4.3.0 through 7.0.1 lets a user who can trigger a Dag, but who cannot author Dags, run arbitrary shell commands on the remote SSH host as the SSH connection's user. SSHRemoteJobOperator quoted templated `environment` values only once even though they were embedded in an already-quoted wrapper script, so payloads such as `$(...)` or backticks were evaluated by the remote shell; the exposed path is a Dag that feeds trigger-time input (`params` or `dag_run.conf`) into `environment`. Exploitation is not possible against a default installation - it requires that non-default Dag design to already exist plus Airflow authorization to trigger that specific Dag - and no public exploit has been identified at time of analysis, with EPSS at 0.27% (20th percentile) and a vuln.today decision label of monitor. The vendor-released fix is version 7.0.1.
Path traversal in Apache SIS 0.7 through 1.6 allows a crafted geospatial data file to make the library open local files referenced through embedded links, regardless of where those files reside. Victims are affected only when they feed untrusted data into one of three parsing paths - GML with xlink:href attributes, WKT with NADCON/NTv2/RGF93 datum-shift grid parameters, or a Landsat scene pointing at per-band GeoTIFF files - and the exposed impact is limited to low confidentiality loss because the target must still match an expected geospatial format and parsing aborts at the first error. No CISA KEV listing or public exploit code was identified at time of analysis, and the vendor rates sensitive-information leakage as unlikely; the fix is Apache SIS 1.7, which confines references to the referencing file's directory or a predetermined grid directory.