SkillPedia Skills Encyclopedia
中文 EN

Skills 2026-08-22 00:25 UTC

llm-sast-scannerlatest

Version [latest] committed by SunWeb3Sec. Update Category: Feature Addition & Enhancement. Commit details: sink-first sweep and under-reporting guardrails. This update introduces new capabilities and module extensions.

SunWeb3Sec fedaf6a SunWeb3Sec/llm-sast-scanner

CommitRaw commit message

sink-first sweep and under-reporting guardrails

Driven by a confirmed miss: two scans of a Spring Boot repo failed to
  find a CWE-22/CWE-73 arbitrary file write that an inline scan caught.
  Root cause was not a missing pattern — `new File(tainted)` was already
  documented in path_traversal_lfi_rfi.md. Four workflow gaps hid it:

  1. Source-first-only tracing. The target had 369 Java files but only
     8 filesystem sinks; forward taint from hundreds of request params
     never converged. Backward from 8 sinks hits it deterministically.
  2. The path traversal reference was framed almost entirely around
     reading files, so a write primitive read as "not exploitable".
  3. The taint passed through a helper named `createFileName` whose
     safe branch was the fall-through and whose early return passed
     user input through untouched — credited as a sanitizer on its name.
  4. The endpoint was admin-authenticated, so trusted-role suppression
     rules dropped it.

  SKILL.md:
  - Step 3 is now bidirectional. Pass A (sink-first sweep) is mandatory
    and runs first, with a per-language sink table. Every sink must be
    dispositioned TAINTED / CONSTANT / UNREACHED and reported — a scan
    that does not state its sink count has not completed Pass A.
  - Added decoy-sanitizer rule: never infer sanitization from a name;
    the supplied-value branch is the attacker's branch.
  - Added cross-layer/@Async continuation rule, including re-derived
    arguments (caller computes a sanitized local, then passes the raw
    source anyway).
  - New Under-Reporting Guardrails section that overrides the existing
    ~30 suppression rules: authority escalation beats trusted-role
    suppression (an admin may export a CSV, not choose its path on
    disk); constrained impact (forced extension, no read-back, no RCE
    chain) is not grounds to suppress; benign-sounding parameter names
    are not grounds to suppress.

  references/path_traversal_lfi_rfi.md:
  - New Arbitrary File Write (CWE-73) section: write-side sink inventory
    across Java/Python/Node/.NET, detection rules (bare relative paths
    are worse not safer; `resolve()`/`Path.Combine` are not sanitizers),
    four-level impact model that does not require an RCE chain, and the
    full worked example.

  Regression: rescanning the target now reports both sinks CONFIRMED with
  a complete 8-sink inventory, and correctly drops a third candidate as a
  false positive with cited evidence.