danielohayon/cve-rl-tasks-pilot100

CVE vulnerable-function location — pilot (100 tasks)

RL environments built with Harbor. Each task gives an agent a repository checked out at the commit before a security fix, plus a bug report describing the symptoms. The agent must locate the function containing the flaw and report its line range.

100 tasks, drawn from CVEs published 2026-09-14 to 2026-09-23.

The task an agent sees

instruction.md holds the bug report, where to write the answer, and the format:

{"file": "relative/path/from/repo/root.ext", "start_line": 120, "end_line": 168}

The repo is at /repo with its git history stripped, and the agent runs with no network — so the fix cannot be read out of a later commit or looked up online.

Reward

Exact file match, then intersection over union on the line range:

Answer	Reward
Exact span	1.0
Partial overlap	0 < r < 1
Wrong file	0.0
Whole file (1–99999)	~0.0004
Missing or malformed	0.0

IoU rather than plain overlap, so a deliberately wide guess scores near zero instead of full marks.

Layout

cve-2026-77405/
├── instruction.md          the prompt, including the bug report
├── task.toml               config, plus ground truth under [metadata]
├── environment/Dockerfile  clones the repo at the pre-fix commit
├── solution/solve.sh       reference answer, for oracle validation
└── tests/
    ├── test.sh             verifier entry point
    └── grade.py            the IoU grader

Ground truth is included in each task.toml:

answer_file = "connection.go"
answer_function = "tlsConfigFromURI"
answer_start_line = 1305
answer_end_line = 1339

Harbor does not pass [metadata] into the container, so the agent never sees it. But anyone reading this dataset can, which makes it unsuitable as a blind benchmark as published.

Repositories are not included

Each environment/Dockerfile clones its repo at the recorded commit during the image build, so a task directory is a few KB of text rather than a full checkout. That keeps all 100 tasks at ~2.3 MB instead of ~8 GB.

The consequence: building an image needs network access and depends on GitHub still serving that commit. A repo that is deleted or force-pushed cannot be rebuilt.

Running one

harbor run -p cve-2026-77405 -a oracle    # expect reward 1.0

How the bug reports were written

CVE descriptions frequently name the vulnerable file and function outright — true of about 1 in 4 candidates — which would hand the agent the answer. So each description was rewritten as the report a user who hit the problem would file: symptoms only, no identifiers.

The writer was given the maintainers' own pull request (title, body and the diff) plus the target's directory, so a vague advisory still yields a report that points at the right area: "the task log view", "external storage backends". A deterministic check then rejects any report naming the target file, its basename, the qualified function name, function names taken from the diff, or the CVE ID. Words the advisory itself already published are exempt, since the report is a rewrite of that text.

All 100 reports pass that check.

Selection

From a larger pool, a CVE qualifies only when its fix maps to exactly one function: one pull request with mapped functions, one (file, function) pair, a resolvable line span, and a named function (anonymous callbacks are excluded). Function lengths run 3 to 476 lines, median 51.

Languages: Go 25, Python 22, C++ 19, TypeScript 12, PHP 12, Java 7, Rust 2, Ruby 1. 72 distinct repositories.

Caveats

• Answers are in the task files (see above).
• Very large functions are easier under IoU — a 476-line target is a generous span.
• Difficulty varies with repo size. Locating a function in a 3,000-file monorepo is a different problem from one in a small library. --clues N in the generator can narrow the search to a list of N candidate files if the zero-hint version proves too hard for a learning signal.