Validation is targeting the exact code paths that matter

LLM-based testing is becoming effective when generation is anchored to a concrete execution target. DiffTestGen uses call graphs, documentation, and coverage feedback to reach modified Python code. Across 463 pull requests, it exposed behavioral differences in 78.2% and averaged 90.7% union coverage.

GapForge applies the same targeting logic to compiler coverage gaps. It infers the program structures and compiler options needed to reach uncovered regions. In 72-hour runs, it exceeded WhiteFox by 24,736 covered GCC lines and 19,798 LLVM lines, and found 12 compiler failures. These results strengthen the recent evidence for executable checks by showing that the location and structure of feedback matter, not merely the presence of a test loop.

Agent control is being specified as repository policy rather than left to ad hoc review. The Agent Governance Manifest links contribution risk, required evidence, accountability, and maintainer gates. In a small controlled evaluation, exact risk-label recovery improved from 15/37 without its materials to 37/38 with them; perceived review support rose from 3.27 to 6.14 on a seven-point scale.

A published library of 85 engineering loops shows a lighter-weight implementation pattern: each workflow defines checks, stopping conditions, and reviewable artifacts. Its examples are concrete, but it provides no aggregate success-rate or baseline evaluation. Together, the sources support growing procedural specificity, while only the manifest offers controlled evidence of review benefits.

Security evidence now treats both instructions and imported code as parts of the trusted computing base. A parser-driven study finds that removing prompt constraints, guards, conditions, or concept bindings can change the likelihood of insecure code from open large language models. The supplied results do not include effect sizes or per-model vulnerability rates, so the magnitude remains uncertain.

Code-Poisoning Property Inference Attacks demonstrates a more direct supply-chain risk: malicious training code can encode private dataset properties for later extraction through label-only queries. The paper reports 100% attack accuracy across four datasets, eight architectures, and 18 properties without reducing model accuracy. Taken together, the studies argue for reviewing generation instructions and executable dependencies before they enter an automated workflow.