SDLI: security debt banks when agents bank progress
Sep 28 Rajput et al.: Security Debt Line Integral couples SAST risk to historical-best test progress — 3.9%/26.2% agree2-positive among working SWE-bench/ProgramBench runs; gowsdl debt arrives in one write and persists.
Rajput, Kabore, Song, Tessa, Malloy, Klein, and Bissyandé notice that coding agents traverse hundreds of intermediate states while security review usually scans only the final artifact. They define the Security Debt Line Integral (SDLI): at each checkpoint, charge a BBCS SAST risk reading (Bandit / Bearer / CodeQL / Semgrep) against the increase in historical-best test pass ratio. Risk after progress plateaus contributes zero; later cleanup does not refund prior charges. Two-tool CWE agreement (agree2) is the stricter channel.
Among working solutions, scanner findings coexist with green tests: 3.9% of 830 passing SWE-bench runs and 26.2% of 80 ProgramBench runs (≥90% official tests) carry non-zero agree2 endpoint SDLI (union rates much higher; excluding three advisory-heavy classes drops ProgramBench agree2 to 6.2%). Same-task SWE-bench siblings can diverge — eight non-zero and one zero on hatch-422. A reconstructed ProgramBench gowsdl trajectory shows the anatomy overnight factories fear: 62 probing actions at zero risk, then one heredoc write that introduces every later-flagged CWE class, then 53 functional refinements that never move the risk reading. MirrorCode checkpoints almost never decrease agree2 risk once it appears. A single repair case study can lower scanner signal while preserving local behavior — and also shows equivalent API rewrites that satisfy detectors without removing capability.
Why it matters for code review
Continuous verification under volume is not only unit-test green. If agents bank functional progress overnight while security debt banks with it, morning humans inherit a progress-weighted exposure curve they never saw forming. Review always-on loops for checkpoint SAST (or equivalent) gated on progress events — not only final CI — and treat “pass rate up, risk flat” as incomplete. Bursts of supervised Claude still win when you cannot afford trajectory monitoring; harnessed factories that skip it compound insecure architecture choices under the same FOMO that celebrates unattended merge rate.