v0.1.2 implementation plan (Track 1 + Track 5)

Goal: Close the roadmap “immediate” slice: deeper extraction on MLflow/Jupyter and any remaining file-discovery gaps, with clear acceptance tests.

References: roadmap.md (v0.1.2 section), pkg/exploit/mlflow/secrets.go, cmd/aipostex/mlflow.go, pkg/exploit/jupyter/secrets.go, cmd/aipostex/jupyter.go, pkg/discover/rules/.

---

1. MLflow: artifact_uri and param connection strings

Current state

• Run.ArtifactURI and Run.Params are already parsed in pkg/exploit/mlflow/client.go.
• pkg/exploit/mlflow/secrets.go ExtractSensitiveParams classifies cloud artifact URIs (S3, GCS, Azure-style paths) and connection-like param values.
• cmd/aipostex/mlflow.go runMLflowRuns already emits per-run info findings and additional High findings for each sensitive match, with evidence on the finding.

Remaining work (if any)

• Audit vs roadmap: Snowflake-style artifact_uri covered by TestExtractSensitiveParams_SnowflakeArtifactURI in pkg/exploit/mlflow/secrets_test.go.
• Credchain: internal/credchain/credchain.go extractMLflowRunHint registers mlflow-run-id; internal/credchain/autochain.go suggests mlflow artifacts --run-id. Tests: TestExtractMLflowRunID, TestGenerateChainActions_MLflowRunID.
• Docs: docs/modules/mlflow.md and roadmap Shipped section updated.

Exit criteria: Roadmap bullet is either marked shipped with test pointers, or a single PR adds the missing scheme/credchain/doc with tests.

---

2. Jupyter: mine notebook cells for secrets

Current state

• pkg/exploit/jupyter/secrets.go MineNotebookSecrets scans notebook JSON cell sources.
• cmd/aipostex/jupyter.go read-notebook fetches content and emits secret findings. jupyter notebooks --mine-secrets lists then parallel-fetches each notebook (worker cap 8, uses cfg.Concurrency).

Implementation tasks

• Option A: --mine-secrets on jupyter notebooks with bounded parallelism.
• Option B: New subcommand jupyter mine-notebooks (not needed; flag covers the use case).
• Reuse newExploitFinding / metadata aligned with read-notebook secret findings (action: notebooks, mine_secrets: true).
• Tests: cmd/aipostex/jupyter_notebooks_test.go TestJupyterNotebooksMineSecretsFindsCredentialInSecondNotebook.

Exit criteria: Operator can obtain cell-level secret findings without manually running read-notebook per path; tests cover at least one multi-notebook list scenario.

---

3. File discovery: model files + training data

Current state

• pkg/discover/rules/local_llm.yaml already includes GGUF, SafeTensors, pickle, PyTorch (.pt/.pth/.bin), ONNX, Ollama paths, Docker AI compose patterns.
• pkg/discover/rules/core_assessment.yaml includes fine-tuning manifests (jsonl/parquet + path hints + content patterns), Arrow, TFRecord, embedding/RAG patterns, ML Training CSV Tables, Hugging Face Dataset Manifest.

Remaining work

• Gap review: CSV rules constrained to dataset/fine-tune/sft/train path globs plus header heuristics.
• Hugging Face / manifest: dataset_infos.json / dataset_dict.json with path patterns.
• max_file_size semantics: pkg/discover/files.go — default 10MB only when content_patterns exist; filename-only rules keep MaxFileSize 0. matchFile applies 10MB default only when compiledContent is non-empty.
• Tests: dedicated fixture test for new YAML lines (optional; go test ./pkg/discover/... passes with embedded rules load).

Exit criteria: Roadmap Track 5 v0.1.2 bullets are either explicitly satisfied by existing rules (document in roadmap “shipped”) or implemented with tests in one PR.

---

4. Suggested sequencing

1. Jupyter notebooks mining (largest behavioral gap).
2. MLflow audit / small extensions + doc touch-up.
3. File rule gaps after a quick inventory against pkg/discover/rules/*.yaml.

---

5. Release checklist

• go test -race -tags=integration ./... (run before release)
• Update roadmap.md Shipped / v0.1.2 section to reflect what actually landed
• Lab coverage target (roadmap: 95%+) re-run if you track it in coverage.md