CuraSec

tag: Ml-Security · 4 items

Act Hugging Face malware campaign returns, hijacks GPUs and plants backdoors
2026-10-30 · BleepingComputer · source ↗ #supply-chain#ml-security#crypto-mining
  • Engineer — Act: Supply-chain compromise in a major ML registry qualifies as Act even without KEV: audit Hugging Face model pull history across pipelines, verify model hashes against known-good sources, and inspect GPU workloads for unauthorized mining processes. Also review any AI-agent-suggested GitHub starter kits for RAT payloads.
  • SOC/IR — Plan: No specific IOCs are provided to sweep for immediately, but the NetSupport RAT delivery via developer toolchains is a detection gap worth closing this quarter — build or tune detections for NetSupport execution originating from CI/CD runners and anomalous GPU utilization spikes on ML infrastructure.
  • Leader — Learn: ML supply chain attacks are maturing from nuisance to persistent threat; share with AI/ML team leads to prompt model provenance verification practices, and consider whether the risk register needs a dedicated ML-dependency entry alongside traditional software supply chain risk.
Learn StratFL: Game-Theoretic Defense Against Covert Federated Learning Poisoning
2026-10-19 · arXiv cs.CR · source ↗ #federated-learning#model-poisoning#ml-security
  • Engineer — Learn: Academic research introducing a game-theoretic defense for covert model poisoning in FL deployments; relevant only if you run federated learning in healthcare or finance pipelines, and no production-ready implementation is available yet.
  • SOC/IR — Skip
  • Leader — Skip
Plan Empirical Study: Vulnerable AI Models Widely Adopted Across OSS Repos
2026-10-12 · arXiv cs.CR · source ↗ #ai-supply-chain#ml-security#open-source
  • Engineer — Plan: If your team loads models from HuggingFace, audit dependencies against InsightDB’s flagged-model list and add provenance verification and model-scanning steps to your ML pipelines; the study found no sanitization logic in any of the observed integration patterns.
  • SOC/IR — Learn: The research maps recurring integration anti-patterns (bypassed safety parameters, delegated trust) that improve understanding of how malicious models could persist post-compromise, but no IOCs or ATT&CK-mapped TTPs are provided for immediate detection work.
  • Leader — Plan: The finding that 22% of publicly flagged vulnerable models are actively referenced across 5,000+ repositories quantifies AI supply chain exposure well enough to drive a policy and tooling review this quarter; use this data to scope an internal audit of third-party model usage before the board asks.
Learn archived Code-Poisoning Property Inference Attack Leaks ML Training Data
2026-07-20 · arXiv cs.CR · source ↗ #ml-security#supply-chain#privacy
  • Engineer — Learn: Novel attack vector where malicious code from public repos or coding agents embeds property-inference backdoors into ML training pipelines — no active exploitation or PoC, but teams training models on sensitive data (PII, clinical records) should factor code provenance auditing into their ML supply chain reviews.
  • SOC/IR — Skip
  • Leader — Learn: Research demonstrates that outsourced or open-source ML training code can be weaponized to leak properties of private training datasets; useful framing for AI governance policies covering code provenance in sensitive ML pipelines, but no immediate action is warranted.