Public AI CVE Selection & Analysis

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R35 — NIST NVD & CISA KEV advisories

   R05 — MITRE ATLAS case studies

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Select 3 relevant public AI/ML framework or MCP security CVEs/advisories (e.g., unsafe deserialization, SSRF in document loaders, path traversal, or BOLA/RCE in open-source AI tooling). Map affected versions, preconditions, and exploit chains.

   Reproduce one selected issue inside a strictly isolated local container or analyze its exact patch diff and published proof-of-concept. *(Never test unowned systems).*

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   No separate task text for this block on day 71 — the sourced work sits in the block above. Keep the time box.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Technical teardown notes for 3 public AI CVEs + 1 local isolated reproduction/patch analysis log.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R05MITRE ATLAS↗ atlas.mitre.org

  ML attack tactics/techniques: prompt injection, data poisoning, model extraction, and exfiltration.

• R35NIST NVD & CISA KEV↗ nvd.nist.gov

  CVE records, CPE affected-version ranges, CVSS v3/v4 scoring context, and Known Exploited Vulnerabilities catalog.

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsMITRE · reference matrix
  MITRE ATLAS™

  Gives the tactic/technique IDs (AML.T0051 prompt injection, AML.T0018 model manipulation, AML.T0020 training data poisoning, AML.T0024 exfiltration via inference API) you need to map a selected CVE's exploit chain to an adversary technique instead of a generic bug class.

  R05checked 2026-11-11
• docsCISA · catalog
  Known Exploited Vulnerabilities Catalog

  Authoritative list of vulnerabilities exploited in the wild — use it to pick a teardown target that actually matters rather than one that merely has a CVE id.

  R05checked 2026-11-12
• videoIBM Technology · 8:40
  Anatomy of an AI ATTACK: MITRE ATLAS

  Shows how a real AI attack decomposes into staged tactics and techniques, which is the reasoning pattern required to turn a CVE record into a multi-step exploit chain with preconditions.

  R05checked 2026-11-11
• articleRAND Corporation · research report
  A Structured Approach to Identifying and Characterizing AI Vulnerabilities

  Draws the CWE-vs-CVE distinction for AI systems and explains why some AI weaknesses are structural rather than patchable — exactly the judgement call when you decide which three advisories are worth selecting.

  R05R35checked 2026-11-11
• repoGitHub (ibondarenko1) · CVE matrix + hardening checklist
  llm-serving-security

  A per-framework CVE matrix (vLLM, Triton, Ollama, TGI, SGLang) with vulnerability class, affected and fixed versions, and severity — the exact table you need to pick three AI-stack CVEs and record their CPE version ranges.

  R35checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.