Skip to content

Research: Autogrep-style patch-to-rule mining for ownership regression candidates (bounded spike) #461

Open

Description

@PhysShell

Classification

P3 / bounded research / no automatic rule adoption or production changes.

Source: LambdaSec Autogrep (automatically proposes Semgrep rules from security-fix patches using LLMs, replays them against vulnerable revisions, then filters candidates).

Hypothesis

Can the patch → candidate general rule → old/new replay → negative controls → reviewed diagnostic/fixture workflow yield one useful new ownership regression case for Own.NET, beyond the existing analyzer tests and verified finding inbox?

Autogrep's actual domain is security Semgrep patterns, not C# lifetime semantics. Do not assume its rules can express disposal, aliases, exceptions, resource-state transitions or interprocedural ownership obligations. Treat it as a methodology donor, not a dependency or a second analyzer.

Scope / method

  • Pin Autogrep commit/license and inspect patch interpretation, rule generation, validation on vulnerable/fixed revisions and dedup filtering. Distinguish the documented pipeline from reproducible guarantees.
  • Inventory existing Own.NET issues and tests for 2–3 confirmed bug fixes (e.g., conditional Dispose, CTS use-after-dispose, handler/event lifetime); avoid unrelated security CVEs.
  • Predefine independent expected results for old bug / fixed code / similar-but-safe negative control / alias or branch mutant, with Roslyn/OwnIR completeness caveats.
  • Try, first on paper, to express the property as: (A) syntax-local Semgrep pattern, (B) OwnIR fact/constraint, or (C) generated fixture only. Measure whether abstraction loses key semantics. If most cases require B/C, reject automated Semgrep adoption.
  • If a minimal experiment is subsequently approved, compare candidate yield, false positives, duplicate rate, manual adjudication cost and unique new failures against existing regression practices. Synthetic/public cases only.
  • KILL if no distinct, independently verified new failure/fixture emerges or manual review exceeds its value.

Related

OWN.NET #405 (semantic mutation), #402 (CTS false-positive inversion), #426 (completeness/evidence), #452 (proof assumptions), and the existing finding-inbox verification workflow. Never auto-publish generated GitHub issues, diagnostics or fixes from inferred patches.

Deliverable

A short KEEP / KILL memo with pinned references, 2–3 example patch→property transformations, exact negative controls, and a decision on whether to reuse only the methodology. This issue does not authorize tests/PRs, workflow automation, new runtime dependencies or CI gating.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions