mirror of
https://github.com/norandom/OpenUltraSAST.git
synced 2026-10-17 12:09:27 +00:00
Marius CiepluchandClaude Fable 5.1
8ab4207618
Tasks 1.1, 5.1, 5.2, 5.4 and the data half of 5.3. - mechanisms.toml: closed vocabulary v1 (20 ids) every expected row cites. - vibe-py: 35 Real-Vuln-Benchmark pairs (Apache-2.0), human provenance only because the LLM-generated repositories carry no license (escalated). - vfc-js: 17 Node fix-commit pairs harvested upstream (MIT/Apache-2.0), SecBench.js itself is not vendored (no license); sinks derived on 13/17. - agent-vfc: 29 agent-trailer fixes as catalog-candidates.toml plus REVIEW.md; the loaded catalog stays empty until a human reviews them. - datasets.toml and README document every slice and its review status. Co-Authored-By: Claude Fable 5.1 <[email protected]> Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG
94 lines
4.7 KiB
Markdown
94 lines
4.7 KiB
Markdown
# Vuln vs fixed pair corpus
|
||
|
||
This corpus measures harness **efficiency**, not cheat-sheet recall.
|
||
|
||
The stage-1 smoke gate (`python -m openultrasast.gate`) scans one tree that
|
||
already contains the planted sinks. A pair eval instead scans **two isolated
|
||
trees**:
|
||
|
||
1. **vuln** — only the vulnerable file. The harness should fire (true positive).
|
||
2. **fixed** — only the patched file, same relative path. The harness should stay
|
||
silent on that CWE (true negative).
|
||
|
||
`pair_correct` = detected on vuln **and** silent on the fix. That is the signal
|
||
the improvement loop consumes (`miss` on the vuln side, `fp` on the fix side).
|
||
|
||
## Layout
|
||
|
||
| Path | Role |
|
||
| --- | --- |
|
||
| `catalog.toml` | Runnable offline pairs (local fixtures + vendored GitHub functions) |
|
||
| `datasets.toml` | Pointers to public VFC datasets — do not vendor the dumps |
|
||
| `datasets/cwe-bench-java-project_info.csv` | 120 vetted Java CVE buggy/fix commits |
|
||
| `github/` | Isolated function excerpts with provenance headers |
|
||
|
||
Pairs in this directory **must not** join `LANGUAGE_MANIFESTS`.
|
||
|
||
## Slices
|
||
|
||
- **local** — in-tree fixture counterparts (`app.py` vs `safe.py`, …). CI gate:
|
||
every local pair must pass (`python -m openultrasast.pair_gate`).
|
||
- **github** — real commit/function pairs (SVEN, CWE-Bench-Java, NVD patterns).
|
||
Measured, not a 90% smoke gate. Misses here are the honesty / evolve-the-harness
|
||
dashboard.
|
||
- **sast** — OWASP Benchmark Java/Python true-vs-false files and Juliet
|
||
bad-vs-good sinks (`benchmarks/pairs/sast/`). Score is Youden (TPR − FPR).
|
||
Honesty dashboard, not a CI fail.
|
||
- **vfc** — reviewed isolated functions from OpenSSL, Firefox, Chromium, and
|
||
curl (`benchmarks/pairs/vfc/`). Overlay vs inventory scoreboard. Honesty
|
||
dashboard, not a CI fail. Training labels (vuln vs fixed) are
|
||
`vfc/training/manifest.jsonl` (vuln vs fixed, one snapshot per line).
|
||
Grow with a blobless `git log --grep=CVE-202` mine plus `vfc/harvest.py`,
|
||
then `vfc/generate_catalog.py` (maintainer, not CI).
|
||
- **vibe-py** — Real-Vuln-Benchmark functions (`benchmarks/pairs/vibe-py/`):
|
||
Python web repos pinned by commit, one vulnerable function per pair with a
|
||
false-positive trap from the same repo as the fixed twin. Today only the 15
|
||
licensed human-authored repos are vendored (`provenance = "human"`); the 40
|
||
LLM-generated repos carry no license file and are excluded until one is
|
||
stated upstream. Educational/CTF apps, not production. Honesty dashboard.
|
||
- **vfc-js** — server-side JavaScript function pairs harvested from the upstream
|
||
fix commits that SecBench.js records (`benchmarks/pairs/vfc-js/`). Command
|
||
injection, path traversal, and code injection first; prototype pollution and
|
||
ReDoS carry their mechanism and may be `known_limit`. No SecBench.js file is
|
||
vendored. Honesty dashboard.
|
||
- **agent-vfc** — security-fix commits carrying an AI agent co-author trailer or
|
||
generation marker, in JavaScript, TypeScript, or Python
|
||
(`benchmarks/pairs/agent-vfc/`). `provenance = "agent"`. Rows enter
|
||
`catalog.toml` only after a human sets `reviewer` in `recipes.toml`; until
|
||
then they sit in `catalog-candidates.toml`, which is never loaded. Honesty dashboard.
|
||
|
||
## Labels, profiles, and mechanisms
|
||
|
||
Every `[[pair.expected]]` row names a `mechanism` from the closed vocabulary in
|
||
`benchmarks/pairs/mechanisms.toml`; every pair carries `provenance`
|
||
(`human|agent|mixed|synthetic`), a declared `split` (`train|holdout`), and an
|
||
optional `known_limit` reason. The scorer counts overlay `coverage` records with
|
||
a source as detections and as leaks, matches by function when a label names one,
|
||
never matches a CWE-only row (`weak_label`), and reports `per_profile`,
|
||
`per_mechanism`, `achievable` (known-limit pairs excluded), and `loss`
|
||
(parse-failed files, unadjudicated proposals). `ousast pairs --profile agent
|
||
--split holdout --json` filters; `--hunter` adds the LLM hunter as a third scorer
|
||
when a hunter model is configured. The improve loop rejects a change that
|
||
regresses any profile on the holdout split.
|
||
|
||
Harvest for every slice goes through `benchmarks/pairs/harvest.py --slice <name>`
|
||
(modes `name`, `line_range`, `hunk`, `enclosing`; comment and string mentions
|
||
never anchor) and `benchmarks/pairs/catalog_gen.py --slice <name>` regenerates
|
||
the catalog from `recipes.toml`. Maintainer only; CI never fetches.
|
||
|
||
## Commands
|
||
|
||
```bash
|
||
uv run ousast pairs
|
||
uv run ousast pairs --slice github --json
|
||
uv run ousast pairs --slice vfc --json
|
||
uv run python -m openultrasast.pair_gate
|
||
```
|
||
|
||
## Growing the corpus
|
||
|
||
Prefer function-level before/after from a reviewed source (SVEN jsonl, a
|
||
CWE-Bench-Java `fix_info.csv` row, a GHSA with a tiny patch). Isolate the
|
||
changed function into `vuln` / `fixed` files with the same `relpath`. Leave
|
||
whole-repo clones to `--fetch` later; CI stays offline.
|