You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
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