Datasets:

quguanni
/

kernel-vuln-dataset

@@ -1,126 +1,168 @@
-# A minimal formal model for LLM tool permissions + a tiny detector for prompt-injection-induced grants
-**TL;DR.** LLM agents call tools. Prompt injection can coerce the agent into **granting new rights** (e.g., `browser.read`) it wasn’t authorized to have — a classic **grant/take/transfer** abuse. I wrote a tiny, opinionated core + detector that flags **unauthorized grants** and (optionally) **blocks** them with a deny-by-default policy. Minimal moving parts, fast to reason about, easy to extend.
 ---
-## Why (problem in two beats)
-- Tool-using agents are great… until the prompt smuggles in a **permission change**.
-- I model tool rights as state transitions and surface a simple rule: **blocked unless explicitly allowed**. If something tries to sneak a grant, the detector yells in red.
 ---
-## Model (small, sharp, composable)
-- **Subjects / Objects / Rights**
-  - Subject = `assistant`
-  - Objects = tools `{browser, file_system}`
-  - Rights = `{use, grant, take, transfer}`
-  - For the demo I instantiate `use.read` and `grant.read`.
-- **State**: `rights_state ⊆ (subject × tool × right)`
-- **Transition**: events update state via `grant` / `use`
-- **Policy**: `allowed_grants = ∅` (deny-by-default) → any `grant` is **unauthorized** unless whitelisted
-Threats I care about first:
-1. **Data exfiltration** via injected `browser.read`
-2. **Tool escalation** (grant write/upload)
-3. **Cross-tool lateral movement** (chained grants to new tools)
----
-## Detector (baseline, intentionally tiny)
-- **Rule**: if an event of type `grant` is **not** in `allowed_grants`, flag
-  `UNAUTHORIZED GRANT DETECTED`.
-- **Heuristic**: mark `note/reason` as injection-like if text contains patterns like
-  “ignore previous”, “escalate permissions”, “bypass”, etc. (cheap, effective baseline).
-**Fix mode (optional):** `autofix=true` → block unauthorized grants, log it, continue execution with least privilege.
----
-## Repo layout
-```
-llm-perm-detector/
-├─ perm_model/
-│  └─ model.yaml          # tools/rights + allowed_grants (default: empty)
-├─ engine/
-│  └─ perm_engine.py      # tiny state machine (grant/use) + deny-by-default
-├─ detector/
-│  └─ detector.py         # injection text heuristic (baseline)
-├─ cases/
-│  ├─ benign.json         # no permission change → no flags
-│  ├─ injected.json       # injection asks for browser.read → flagged
-│  └─ injected_autofix.json # same as above but blocked by autofix
-└─ simulate.py            # glue runner; pretty prints log + summary
 ```
----
-## Quickstart
-```bash
-# benign (should be clean)
-python3 simulate.py cases/benign.json
-# injected (should flag an unauthorized grant)
-python3 simulate.py cases/injected.json
-# injected + autofix (should flag AND block)
-python3 simulate.py cases/injected_autofix.json
-```
-Expected money lines:
-- Benign:
-  `No unauthorized grants detected.`
-- Injected:
-  `UNAUTHORIZED GRANT DETECTED: assistant -> browser.read  reason='...'`
-- Injected + Autofix:
-  `[DETECTOR] UNAUTHORIZED GRANT DETECTED: ...`
-  `[AUTOFIX] Blocked unauthorized grant.`
-**Screenshot the three runs** (benign, injected, injected+autofix). That’s the demo.
----
-## Policy knob (optional whitelist)
-You can explicitly allow a grant in `perm_model/model.yaml`:
-```yaml
-policies:
-  allowed_grants:
-    - ["assistant", "file_system", "read"]
 ```
-Re-run to see the contrast: whitelisted grants flow; everything else gets flagged/blocked.
----
-## Tiny benchmark plan (what I’m targeting)
-- **40 prompts** total (20 benign / 20 injected)
-- KPI goal by **Nov 6**: **precision ≥ 0.80**, **recall ≥ 0.70** on unauthorized grants
-- For the demo here, I show the 3 canonical cases above
----
-## Roadmap (2 weeks)
-- **v0.1**: formal core + detector + 40-prompt micro-benchmark; **min-fix playbook**; **5-page draft** + tidy repo
-- **Stretch**: add `transfer` events; a simple lattice for right inclusion; evaluate on a more diverse injection set; logs → JSONL for downstream dashboards
----
-## Impact (why this framing)
-- Turns “prompt injection” into concrete, auditable **state transitions**.
-- Makes agent tool-use **observable** and **enforceable**.
-- Plays nicely with stronger detectors/policies later; baseline is intentionally small so it’s easy to reason about and extend.
----
-If you want to skim one thing, run the three cases and look for the red line. If you want to extend it, add a new tool/right in `model.yaml`, drop a new `cases/*.json`, and watch the state machine make the implications obvious.

 ---
+license: mit
+task_categories:
+  - text-classification
+language:
+  - en
+tags:
+  - linux-kernel
+  - security
+  - vulnerabilities
+  - code
+  - bugs
+  - software-engineering
+pretty_name: Linux Kernel Vulnerability Dataset
+size_categories:
+  - 100K<n<1M
 ---
+# Linux Kernel Vulnerability Dataset
+**125,183 bug-introducing commits** mined from 20 years of Linux kernel git history (2005-2025).
+## Dataset Description
+This dataset maps every kernel bug fix with a `Fixes:` tag back to the commit that introduced the bug. It enables research on:
+- **Vulnerability detection**: Train models to identify bug-introducing commits
+- **Bug lifetime analysis**: Study how long different bug types persist
+- **Subsystem risk assessment**: Identify which kernel subsystems have the longest-lived bugs
+- **Developer patterns**: Analyze who introduces and fixes bugs
+### Key Statistics
+| Metric | Value |
+|--------|-------|
+| Total bug-fix pairs | 125,183 |
+| Valid records (0 < lifetime < 27 years) | 123,696 |
+| Average bug lifetime | 2.1 years |
+| Median bug lifetime | 0.7 years |
+| Longest-lived bug | 20.7 years |
+| Bugs hiding 5+ years | 13.5% |
+| Unique fixing commits | 119,449 |
+| Unique bug-introducing authors | 9,159 |
+| With CVE ID | 158 |
+| With Cc: stable | 27,875 (22%) |
+## Dataset Structure
+### Data Fields
+| Column | Type | Description |
+|--------|------|-------------|
+| `fixing_commit` | string | SHA of the commit that fixed the bug |
+| `introducing_commit` | string | SHA of the commit that introduced the bug |
+| `lifetime_days` | int | Days between introduction and fix |
+| `subsystem` | string | Kernel subsystem (networking, drivers, fs, etc.) |
+| `bug_type` | string | Detected bug type (race-condition, null-deref, etc.) |
+| `fix_subject` | string | Commit message subject of the fix |
+| `fixing_date` | datetime | When the fix was committed |
+| `introducing_date` | datetime | When the bug was introduced |
+| `files_changed` | string | Files modified in introducing commit |
+| `insertions` | int | Lines added in introducing commit |
+| `deletions` | int | Lines removed in introducing commit |
+| `cve_id` | string | CVE ID if mentioned in fix commit (nullable) |
+| `cc_stable` | bool | Whether fix was tagged for stable backport |
+| `fix_author` | string | Author of the fix |
+| `intro_author` | string | Author who introduced the bug |
+### Subsystem Distribution
+| Subsystem | Count | Avg Lifetime |
+|-----------|-------|--------------|
+| drivers | 38,412 | 2.3 years |
+| networking | 6,079 | 2.9 years |
+| gpu | 5,212 | 1.4 years |
+| filesystem | 4,891 | 2.6 years |
+| sound | 3,847 | 2.1 years |
+| usb | 2,505 | 3.5 years |
+| memory | 2,459 | 1.8 years |
+| bpf | 959 | 1.1 years |
+### Bug Type Distribution
+| Bug Type | Count | Avg Lifetime |
+|----------|-------|--------------|
+| null-deref | 4,931 | 2.2 years |
+| use-after-free | 2,963 | 3.2 years |
+| memory-leak | 2,846 | 3.1 years |
+| refcount | 2,209 | 2.8 years |
+| deadlock | 1,683 | 2.2 years |
+| race-condition | 1,188 | 5.1 years |
+| buffer-overflow | 399 | 3.1 years |
+| integer-overflow | 298 | 3.9 years |
+## Collection Methodology
+### Mining Process
+1. Clone Linux kernel repository at v6.19-rc3
+2. Run `git log --grep="Fixes:"` to find all fixing commits
+3. Extract the referenced commit hash from each `Fixes:` tag using regex
+4. Retrieve metadata (dates, authors, files) from both commits
+5. Classify subsystem from file paths using 70+ regex patterns
+6. Detect bug type from commit message keywords
+7. Calculate lifetime as `fixing_date - introducing_date`
+### Extraction Code
+```python
+import re
+def extract_fixes_tag(commit_msg: str) -> str | None:
+    """Extract the commit ID from a Fixes: tag"""
+    pattern = r'Fixes:\s*([a-f0-9]{12,40})'
+    match = re.search(pattern, commit_msg, re.IGNORECASE)
+    return match.group(1) if match else None
 ```
+### Filtering
+- Excluded commits where `Fixes:` tag points to non-existent hash
+- Excluded merge commits
+- For lifetime analysis: excluded same-day fixes (lifetime_days = 0)
+- For lifetime analysis: excluded outliers (lifetime > 10,000 days)
+## Limitations
+- **Selection bias**: Only ~28% of fix commits use `Fixes:` tags. Well-documented bugs (likely more serious) are overrepresented.
+- **Mainline only**: Doesn't include bugs fixed only in stable branches or vendor kernels.
+- **Subsystem classification**: Heuristic-based on file paths; some bugs span multiple subsystems.
+- **Bug type detection**: Keyword-based; many bugs classified as "unknown".
+- **Timestamp accuracy**: Uses author dates, not commit dates. Rebasing can skew.
+## Intended Uses
+### Recommended
+- Training vulnerability detection models (e.g., VulnBERT)
+- Analyzing bug lifetime patterns across subsystems
+- Studying the evolution of kernel security over time
+- Research on code review and testing effectiveness
+### Out of Scope
+- Direct exploit development (commits don't include full code context)
+- Claiming completeness of kernel vulnerability data
+## Citation
+```bibtex
+@dataset{qu2026kernelvuln,
+  author = {Qu, Jenny Guanni},
+  title = {Linux Kernel Vulnerability Dataset},
+  year = {2026},
+  publisher = {HuggingFace},
+  url = {https://huggingface.co/datasets/quguanni/kernel-vuln-dataset}
+}
 ```
+## License
+MIT License
+## Contact
+- **Author**: Jenny Guanni Qu
+- **Email**: jenny@pebblebed.com
+- **Blog post**: [Kernel bugs hide for 2 years on average](https://pebblebed.com/blog/kernel-bugs)