⭐ Star on GitHub

Summary: Wiki pages for raw/Supply-chain-attacks/ — the AI supply chain taxonomy and its attacks, plus (filed here but topically adjacent) securing LLM applications and AI coding assistants.

Sources: raw/Supply-chain-attacks/ (8 files, ~39,400 words)

Last updated: 2026-07-29


Pages

PageCovers
ai-supply-chainThe five-layer taxonomy, grey zones, what differs from traditional software
supply-chain-attacksFive attack paths, with real incidents
provenance-and-assuranceSLSA, SBOM, in-toto, signing — the defensive centre of gravity
do-users-write-insecure-codeStanford CCS ‘23 — the vault’s best empirical evidence on AI coding risk
ai-assistant-instructionsOpenSSF guidance on instruction files as a security control
securing-llm-applicationsDefences for code that calls models — a previously uncovered layer

Two primary sources worth knowing

Plexal & Oxford GCSCC, Securing the AI Supply Chain (March 2026), supported by LASR — the first non-vendor institutional supply-chain source in the vault. Workshop-derived, and the origin of the five-layer taxonomy.

Perry, Srivastava, Kumar & Boneh (Stanford), Do Users Write More Insecure Code with AI Assistants?, ACM CCS ‘23 — peer-reviewed controlled study, 47 participants, five tasks, three languages.

Between them these are the two most citable sources added to the vault in the last three ingests.

Five findings

  1. 78% of organisations have limited or partial visibility of their own AI supply chain; 28% had changed vendors over trust or compliance concerns (Plexal/Oxford).
  2. Developers with AI assistants wrote less secure code — and were more confident it was secure. The tool inverts the calibration between confidence and correctness. See do-users-write-insecure-code.
  3. The CIA triad is insufficient for AI supply chain risk. Plexal/Oxford add explainability, fairness, robustness, reproducibility and accountability — reaching independently a list that overlaps NIST’s seven characteristics.
  4. Provenance beats detection when the artifact isn’t inspectable. Google’s argument for models mirrors CMU’s for training data (chain of custody) — two domains, one conclusion.
  5. model.load() is a code-execution sink. Wiz’s s1ngularity case and OWASP’s “malicious pickling” describe the same mechanism from two directions.

Real incidents documented here

  • Ultralytics — widely-used AI library compromised via GitHub, shipping cryptominers into training and inference environments
  • s1ngularity — malicious payloads embedded in trusted model artifacts, “turning the model loading process itself into an execution vector”
  • CVE-2025-23266 “NVIDIAScape” — privileged GPU-runtime vulnerability exposing shared workloads

A scope note

The folder name says supply chain; three of the eight sources are about AI coding assistants and LLM application security instead. Pages follow the raw-folder convention, but topically:

All are cross-linked both ways. Full note in log.

Gaps this closes, and one it opens

Closed: the vault’s last empty source folder; the “secure patterns for building LLM applications” gap; and the absence of vendor-independent evidence on AI coding risk.

Opened: multimodal prompt injection — instructions embedded in images that vision models interpret and follow — has no PITAX technique and no ATLAS technique. Nearest neighbours are T26 Spatial Byte Arrays and T03 Binary Streams, neither of which covers vision-model instruction-following.

Log

See log.