Training Data
Training data is both the model's most valuable input and its most underprotected one. Three problem classes dominate. First, poisoning: an attacker who can influence a public dataset, a web crawl, or a fine-tuning corpus can plant backdoors or biases that survive into the deployed model — BadNets-style attacks on image classifiers, trigger-phrase attacks on LLMs, and reward-hacking on RLHF datasets. Second, memorization and leakage: models can regurgitate verbatim training data, exposing PII and copyrighted content; this has driven the active New York Times v. OpenAI litigation and is a recurring GDPR concern. Third, provenance: when training data origins are unclear, downstream users inherit legal and security risk they can't assess. EU AI Act Article 10 (Data Governance) and ISO 42001 Annex A treat training-data quality as a controlled asset. Defenses: data lineage tracking, deduplication, PII scrubbing before training, and adversarial training against known trigger families.
| Severity | CVE | Headline | Package | CVSS |
|---|---|---|---|---|
| MEDIUM | CVE-2026-69146 | MLflow: missing authZ lets users forge dataset lineage | mlflow | 6.5 |
| HIGH | GHSA-wg9g-w2j2-8pgr | MONAI: NumpyReader RCE via forced allow_pickle=True | monai | 7.8 |
| HIGH | GHSA-rghg-q7wp-9767 | MONAI: OS command injection via crafted YAML config | MONAI | - |
| HIGH | GHSA-qxq5-qhx6-94qw | MONAI: unpatched pickle deserialization enables RCE | monai | 7.8 |
| LOW | CVE-2026-49431 | FreeBSD ZFS: unprivileged user can spoof recv metadata flag | 3.3 | |
| HIGH | GHSA-p77j-g7h5-r2vw | GeoLens: authz flaws leak private geospatial data | geolens | - |
| HIGH | CVE-2026-76336 | Splunk SPL2: broken authZ lets low-priv user wipe modules | 7.1 | |
| HIGH | CVE-2026-76254 | Splunk Enterprise: SPL injection via Dataset Explorer | 7.5 | |
| MEDIUM | CVE-2026-76341 | Splunk Table Editor: stored SPL escalates power→admin | 5.4 | |
| UNKNOWN | CVE-2026-18286 | aeon: Code injection RCE in activity dataset loader | - | |
| HIGH | CVE-2026-76073 | Label Studio: IDOR exposes cross-org annotations | label-studio | 8.8 |
| LOW | CVE-2026-9805 | InsydeH2O: SMM buffer overflow in IHISI FMTS handler | InsydeH2O | 2.7 |
| UNKNOWN | CVE-2026-69129 | KubePi: broken cluster-scope check exposes MLOps clusters | - | |
| MEDIUM | CVE-2026-82290 | Chainlit: IDOR lets users tamper with others' feedback | chainlit | 5.3 |
| MEDIUM | CVE-2026-7680 | COCO Annotator: path traversal exposes host files | 4.3 | |
| HIGH | CVE-2026-85179 | Label Studio: SSRF via webhooks hits cloud metadata | label-studio | 8.5 |
| HIGH | CVE-2026-85211 | Label Studio: org filter bypass leaks tenant storage | label-studio | 7.7 |
| UNKNOWN | CVE-2026-85063 | csv-parse: CSV header triggers prototype pollution | - | |
| MEDIUM | CVE-2026-85670 | tokenizers: malformed vocab crashes process (DoS) | tokenizers | 6.5 |
| MEDIUM | CVE-2026-70626 | NLTK: symlink escape allows arbitrary file read | nltk | 6.2 |