Security audit
Security checks for vulnerabilities and agentic risk
Optional AI review sends more unpublished research content than the privacy documentation discloses, so this skill needs review before use.
Keep confidential or unpublished work in the local workflow unless you authorize the broader AI upload. Review can send notes, metadata and figure text; --vision additionally sends an image. Concept prompts can include research notes. Use a dedicated, limited-balance key if enabling paid features. Export only trusted SVGs, or use an isolated environment because Chrome's sandbox may be disabled automatically. Independently check journal requirements and AI disclosures. This was a static review; gateway retention and downstream handling were not verified.
scripts/ai.mjsDefault AI review uploads undisclosed research notes and figure text
Unpublished research notes and figure labels, including result values when present as SVG text, leave the machine despite not being included in the stated upload scope. They are sent to each review member rather than remaining local.
Referenced artifact was not completely inspected
所有命令都长这样：`node scripts/fig.mjs <cmd> <工作目录> [参数]`。正常结果走 stdout 的 JSON，进度和人话提示走 stderr，失败返回 `{ok:false,error,fix}` 且退出码非 0。
Referenced artifact was not completely inspected
| `init` | 建工作区：`brief.md` `figure.json` `data/` `out/` | `node scripts/fig.mjs init ./fig1` |
Referenced artifact was not completely inspected
| `spec` | 把 `brief.md` 里的大白话解析成结构化 `figure.json`（本地规则，不调模型） | `node scripts/fig.mjs spec ./fig1` |
Referenced artifact was not completely inspected
| `chart` | `data/*.csv` + figure.json 里的 data 段 → 期刊风格图表 SVG | `node scripts/fig.mjs chart ./fig1` |
Referenced artifact was not completely inspected
/<name>.svg`（pipeline/architecture/mechanism/comparison/concept/toc 排版） | `node scripts/fig.mjs render ./fig1 --journal nature` |
Referenced artifact was not completely inspected
| `check` | 确定性合规检查 → `out/check.json` + 人话报告（每条带 fix） | `node scripts/fig.mjs check ./fig1 --journal ieee` |
Referenced artifact was not completely inspected
| `export` | 按期刊规格导出 PNG/PDF/TIFF | `node scripts/fig.mjs export ./fig1 --formats png,pdf,tif --dpi 600` |
Referenced artifact was not completely inspected
| `concept` | AI 概念示意图/封面/TOC 草稿（默认 `gpt-image-2.5-1k`） | `node scripts/fig.mjs concept ./fig1 --n 2` |
Referenced artifact was not completely inspected
| `review` | 多 AI 会审 → `out/review.json`（**先给修改清单**，分数仅参考） | `node scripts/fig.mjs review ./fig1 --tier econ` |
Referenced artifact was not completely inspected
| `wallet` | 看密钥来源、余额、充值链接（只读，不花钱） | `node scripts/fig.mjs wallet` |
Referenced artifact was not completely inspected
node scripts/fig.mjs wallet # 会告诉你缺 key 和下一步
Referenced artifact was not completely inspected
| 付费接口报 `Invalid token`，但 key 明明是刚复制的 | **余额为 0**，不是密钥坏了（网关就这样回） | `node scripts/fig.mjs wallet` 看余额 → 去充值页充 ¥1 再试 |
Referenced artifact was not completely inspected
想先看"AI 图会错成什么样"，直接看 `references/prompt-library.md` 的「AI 图必错的三件事」和 `references/failure-modes.md` 的样例。
Referenced artifact was not completely inspected
- `references/failure-modes.md` —— 被打回的图长什么样：症状 → 原因 → 改法 → 对应命令
The review workflow sends figure captions, metadata, a serialized portion of spec.data, SVG text, and check results to a remote AI gateway; enabling vision also uploads the figure PNG. Although remote AI review is advertised, the assurance that paper data is not uploaded is misleading because these inputs can contain unpublished results or confidential research, and this module provides no disclosure-and-consent gate before transmission.
Without declared permissions the skill's intent is opaque and cannot be validated.
The manifest claims journal requirements have each been checked against official author guidelines and carry verification dates. In assets/journals.json, L075–L085 instead mark Science's widths, height, minimum font size, DPI and formats as unverified; the same pattern appears for several other publishers. These presets therefore do not support the advertised blanket verification claim, even though their uncertainty is disclosed internally.
The checker asserts that the caption lacks AI disclosure based only on AI-marker presence and figure type, without reading any caption. It can therefore flag properly disclosed figures while skipping undisclosed concept/TOC figures, making the publication-compliance report unreliable. This is especially consequential for a skill marketed as enforcing AI labeling and submission requirements, although it does not establish code execution or data exfiltration.
The documentation at L0692–L0693 explicitly claims that obstacle-aware routing guarantees arrows will not pass through nodes. However, L0712 returns a straight horizontal edge whenever endpoint centers align, without checking intervening obstacles; an edge that skips an intermediate node can therefore pass directly through it, violating a stated diagram-quality guarantee.
L125 specifies that concept figures contain 3–8 English words, and L147 similarly limits TOC phrases to six English words or no text. These prescriptions constrain the output language without user opt-in or a documented, journal-specific language requirement.
The comment explicitly specifies producing a Chinese image-generation prompt, and the default builder uses Chinese instruction templates regardless of the user's language. Although callers can supply a complete custom prompt, the interface does not explicitly offer a language choice or justify the fixed-language default.