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 |
|---|---|---|---|---|
| HIGH | CVE-2026-95619 | libstdc++: integer overflow in aligned new risks corruption | rhoai/odh-mlflow-operator-rhel9 | 7.7 |
| LOW | CVE-2026-15028 | libarchive: PAX SUN.holesdata heap overflow in tar parsing | mlflow | 3.9 |
| HIGH | CVE-2026-84837 | RPM: command injection via tarball name in rpmbuild | rhoai/odh-mlflow-operator-rhel9 | 7.8 |
| MEDIUM | CVE-2026-4426 | libarchive: zisofs field validation flaw enables DoS | rhoai/odh-mlflow-operator-rhel9 | 6.5 |
| LOW | CVE-2025-5916 | libarchive: WARC integer overflow causes local DoS | rhoai/odh-mlflow-operator-rhel9 | 3.9 |
| HIGH | CVE-2026-125214 | FFmpeg: heap overflow in Blu-ray PCM decoder (big-endian) | 7.0 | |
| HIGH | CVE-2026-130584 | Xibo CMS: IDOR in DataSet RSS leaks other tenants' data | 8.3 |
Page 15 of 15