Content
68%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A lean, highly actionable injection payload reference whose code and PoCs are near copy-paste ready. Its weaknesses are a merely implicit detect-then-exploit workflow with no verification checkpoints, and a couple of incomplete or mislabeled payload examples.
Suggestions
Add an explicit detection-to-exploitation sequence with verification steps, e.g., confirm the injected XPath predicate changed the response, then verify document() output before escalating to EXSLT RCE, with fallback engines to try when extensions are disabled.
Complete the blind-extraction section: give a concrete boolean-extraction loop example and name which implementations support timing or out-of-band channels, instead of the current one-line comment.
Fix the EXSLT example attribution (the rt:exec runtime extension is Xalan-Java, not libxslt) and label each RCE payload with the engine and version it applies to so payloads are not fired against the wrong target.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is a dense, payload-first reference with almost no padding; every section delivers attack-relevant material. It is not 5 because a few orientation lines explain concepts Claude already knows ('XPath = XML query language', the SQL-injection analogy), which could be trimmed. | 4 / 5 |
Actionability | Most guidance is copy-paste ready: concrete injection payloads (' or '1'='1), a complete executable curl PoC with a full stylesheet for document() file read, and specific EXSLT RCE payloads. Minor gaps keep it below 5 — the time-based blind XPath technique is only a comment ('XPath has no sleep, but some impls support extensions'), and the EXSLT example mislabels a Xalan-Java extension as 'libxslt w/ Xalan-Java extension'. | 4 / 5 |
Workflow Clarity | Sections (Detection, PoC, Severity, Defender) imply a loose detect-then-exploit flow, but there is no explicit sequence and no validation checkpoints — e.g., confirming document() actually returned content, or falling back to alternate engines when EXSLT is unavailable. The simple-skill exception does not apply (82 lines, multiple attack classes), and the implicit-only sequence matches anchor 3 rather than the explicit-checkpoint anchors 4-5. | 3 / 5 |
Progressive Disclosure | A single self-contained file with clean section headers, appropriate for a compact payload cheat-sheet, plus a cross-references section pointing one level out (upstream corpus, XXE skill) rather than nesting. It does not reach 5 because the body exceeds the under-50-lines threshold where inline-only structure is automatically ideal — engine-specific EXSLT payload tables or blind-extraction technique detail could live in a reference file. | 4 / 5 |
Total | 15 / 20 Passed |