brotli-checker — скилл для Claude Code: что делает и как установить
brotli-checker — это скилл для Claude Code, который ищет уязвимости CVE-2026-111497 в библиотеках Brotli. Инструмент анализирует зависимости и бинарные файлы без запуска кода проекта.
MIT ★ 70 обновлён 2026-11-06
npx skills add trailofbits/brotli-checker --skill brotli-checker -a claude-codeЧто делает brotli-checker
Скилл находит уязвимые версии Brotli внутри исходного кода, архивов и скомпилированных файлов форматов ELF32/64. Он собирает технические данные и переводит их в рекомендации по безопасности. Для работы требуется установленный Python версий 3.12–3.15.
Инструмент сканирует ZIP, JAR, WAR, TAR и другие контейнеры без извлечения содержимого или выполнения команд приложения. При обнаружении уязвимостей скилл отчитывается о путях к файлам и предлагает шаги для исправления. Анализ ELF-файлов использует сигнатуры компиляторов, чтобы проверить патчи в различных сборках.
Какие задачи решает
- Найти уязвимые версии Brotli в проекте
- Проверить скомпилированные бинарные файлы на наличие CVE-2026-111497
- Получить рекомендации по обновлению библиотек
- Проанализировать архивы без выполнения кода
Примеры запросов агенту
- Проверь этот проект на уязвимость Brotli CVE-2026-111497
- Просканируй текущую папку на наличие проблемных версий Brotli
- Какие шаги предпринять для обновления Brotli в приложении?
Как установить brotli-checker
1Через Skills CLI (Vercel)
npx skills add trailofbits/brotli-checker --skill brotli-checker -a claude-code2Вручную: положить файлы
~/.claude/skills/brotli-checker/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/brotli-checker/
1Через Skills CLI (Vercel)
npx skills add trailofbits/brotli-checker --skill brotli-checker -a codex2Вручную: положить файлы
.agents/skills/brotli-checker/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add trailofbits/brotli-checker --skill brotli-checker -a cursor2Вручную: положить файлы
.cursor/skills/brotli-checker/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install trailofbits/brotli-checker skills/brotli-checker2Вручную: положить файлы
.github/skills/brotli-checker/SKILL.md1Через Skills CLI (Vercel)
npx skills add trailofbits/brotli-checker --skill brotli-checker -a gemini-cli2Командой платформы
gemini skills install https://github.com/trailofbits/brotli-checker --path skills/brotli-checkerВручную: положить файлы
.devin/skills/brotli-checker/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о brotli-checker
Как установить brotli-checker в Claude Code?
Выполните команду npx skills add trailofbits/brotli-checker --skill brotli-checker -a claude-code. Инструмент будет доступен по пути ~/.claude/skills/brotli-checker/SKILL.md.
Что делает brotli-checker?
Он автоматически сканирует проект на предмет использования уязвимых версий Brotli, связанных с CVE-2026-111497, и помогает разработчику понять, какие части системы требуют обновления.
Что нужно для работы brotli-checker?
Для полноценной работы скилла на системе должен быть установлен Python версий 3.12–3.15.
Исходный текст
SKILL.md · MIT
--- name: brotli-checker description: Check whether an app or project needs a Brotli security update for CVE-2026-111497, explain the result in plain language, and give practical next steps. Find bundled copies and investigate whether the running app uses them. --- # Check whether an app needs a Brotli update Help someone who is not a security expert answer: "Does my app need an update, and what should I do next?" They should be able to say "check this project" or give a website address without knowing library versions, build details, or the CVE number. Gather technical evidence yourself; translate it into a short, actionable answer. Assess CVE-2026-111497 for the requested source tree, application, or deployment. Finding vulnerable source establishes source containment. Build selection, shipment, attacker-controlled input reaching that decoder, and exploitability are separate conclusions. Keep a result for each copy; one fixed copy cannot clear other copies. ## Gather evidence For "this project", start with the current project folder and discover its application, dependencies, and deployment files. Use the supplied folder or server access when available. Collect versions and package information with trusted, read-only inspection tools; do not run application code or scripts from the target. Do not make the user supply technical facts you can inspect. Identify the target revision, build/platform, deployed artifact or container digest, and the installation actually serving the application when available. If only a URL is supplied, explain that headers and Brotli response support cannot identify the decoder or its patch status. Say "I can't confirm this from the website address alone" and give a short message the user can send to their hosting provider asking it to check and update the app's Brotli decoder. If the conversation already provides local/server access, use it. Ask only for missing access or a target that cannot be inferred. Do not send malformed streams or exploit probes. From this skill's directory, run the offline inventory helper (Python 3.12–3.15). Use JSON internally for the evidence; the default text output is for people: ```sh python3 scripts/brotli_inventory.py /absolute/target --json ``` It reads files and bounded ZIP/JAR/WAR/wheel, TAR, gzip, bzip2, and XZ containers without extracting them or executing target code. Known signatures identify renamed archives and extensionless layers. Recognized unsupported formats such as Debian, RPM, and Zstandard produce coverage gaps; provide unpacked files. Signature-free or unrecognized encodings can be missed. Container traversal does not reconstruct a final image filesystem. It checks eligible file bytes against source fingerprints, identifies modified C decoder patterns, and records dependency, native/Wasm, and JavaScript/TypeScript fetch/Undici candidates. Fetch references include source locations; they are heuristic leads, not confirmed Node calls or attacker-controlled inputs. For supported ELF32/ELF64 targets it also matches compiled decoder text, including stripped and statically linked code. Dependencies, build outputs, and hidden directories are included; `.git` is excluded. Read its coverage gaps. It does not scan an entire host, resolve remote dependencies, follow symlinks, prove whole-application patch status, or establish an absence of Brotli. Supply extracted image filesystems or additional artifact roots as needed. Filesystem symlink targets visited directly do not add a gap; other targets still need checking. If a resource limit is reached, scan a smaller relevant folder or adjust the limit shown by `--help`. Regular files have no default per-file or total byte cap. ELF analysis uses sparse offset reads; full-artifact SHA-256 is streamed. Archive expansion is separately budgeted at 8 GiB/member or stream and 64 GiB total; override with `--max-expanded-bytes` and `--max-total-expanded-bytes` as needed. Optional `--max-file-bytes` and `--max-total-bytes` still impose explicit input caps. For independently obtained upstream versions, optional flags are: ```sh python3 scripts/brotli_inventory.py --brotli-version 1.2.0 --json python3 scripts/brotli_inventory.py --node-version 24.22.0 --json ``` These flags assess a supplied version string, not an installed or running build. Never pass a wrapper package's version as the embedded Brotli version. Version warnings allow vendor backports; prereleases, vendor suffixes, and branches outside the snapshot's rules remain unknown. Exit 1 means affected-source or version evidence; 3 means review/unknown; 0 means only the reported items meet a known source fingerprint or upstream fixed-version rule; 2 means CLI error. The helper never certifies the application safe. For ELF code findings, inspect `evidence.code_matches`: source/object hashes, compiler/flags, state, and file offset. `vulnerable_code_match` and `mixed_code_matches` warrant update review (exit 1). `fixed_code_match` concerns only matched decoder code, leaves other copies unresolved, and does not by itself produce exit 0. Version strings/imports do not determine code state. Matching excludes relocated address fields and does not verify their targets or referenced data or associate separately matched executable sections. Packaged references cover x86, x86-64, ARM/Thumb, AArch64, RISC-V64, PowerPC64 little-endian, s390x, MIPS32/64 in both byte orders, and LoongArch64. The 432 references cover `-O2`/`-O3` PIC/non-PIC builds of three affected/fixed release pairs with GCC 14.2, the original AArch64 GCC 16.2.1 builds, and Clang 19 LoongArch builds. LoongArch also has GCC 14.2 references built with `-mno-relax` against a real LoongArch glibc sysroot. The two Loong toolchains have distinct header/sysroot provenance; inspect `build_note`, `headers_sha256`, compiler, and flags rather than treating the ISA as the build identity. For x86-64 GCC 14.2, AArch64 GCC 14.2, and AArch64 GCC 16.2, the catalog also covers `-ffunction-sections -fdata-sections` linked with section garbage collection. These entries have `code_scope: patch_bearing_sections_v1` and method `relocation_masked_patch_sections`. They require exact matches for every emitted section owning literal, command, and distance block-switch handling, all safe variants, and normal/safe command processing, plus recorded emitted helpers. They intentionally do not require unrelated public APIs which the linker may discard. A fixed scoped match establishes only the recorded patch-bearing population, not the whole decoder, unused APIs, another copy, or runtime selection. Missing any required role prevents that reference from matching. Paired controls reject opposite-state code, and matching both states reports a mixed result. This does not establish that separately matched fragments belong to one decoder copy. Other compilers/options require corresponding trusted references via `--code-references`. Cross-translation-unit LTO can rewrite and reorder the final code beyond both complete and function-section pre-link identities; keep those binaries unknown unless reviewed post-link evidence or bounded disassembly closes the gap. No match remains unknown; a supported target with no references produces an explicit coverage gap. Do not present this as compiler-independent semantic or CFG patch verification. The inventory also reads bounded ELF symbol tables without executing the target. It distinguishes decoder API definitions from imports. For an exactly defined `BrotliDecoderVersion`, it accepts only supported straight-line constant-return instruction forms. A reported old supported-branch version is `potentially_vulnerable` and produces the normal exit 1 unless stronger evidence already applies. A reported fixed version is metadata only and never clears the binary. Code-reference matches remain code-identity evidence; metadata does not upgrade an unmatched binary to `fixed_code_match`. For `decoder_import_candidate`, statically inspect the importer and every in-scope provider candidate before triage. Use a trusted `readelf` to record `DT_NEEDED`, `DT_RPATH`/`DT_RUNPATH`, and undefined decoder symbols, then scan the actual regular provider files from the supplied application, image, or deployment root. Never run `ldd`, invoke a dynamic loader, or shell-evaluate `$ORIGIN` or other path tokens. A needed name, RUNPATH, SONAME, host library, or one fixed provider does not prove which file the deployed loader selects. Keep selection and reachability unknown until deployment-specific evidence closes them. See [dynamic ELF providers](references/reachability.md#dynamically-linked-elf-providers). For `decoder_data_signature`, use the [stable signature guide](references/binary-signatures.md). The full command table or complete code-length table set is stronger recognition evidence than a string, but occurs in both affected and fixed builds. Report `detection_status` separately from patch state. Preserve file offsets, ELF virtual addresses, and mapping scope; locate code cross-references to these **data** tables, then inspect bounded executable ranges and error propagation. Do not disassemble table bytes as if they were code or promote a table match to an affected/fixed verdict. PE/Mach-O/WASM table evidence is not loader validated. Dictionary-only presence does not establish a decoder, and unmatched signatures do not prove absence. For an unresolved local ELF with decoder table, name, symbol, version, or code evidence, collect bounded disassembly for model review instead of stopping at a generic unknown. Do not disassemble every unrelated ELF. Run this only with a trusted local `objdump`; the helper rejects target symlinks, streams the original artifact hash, reads ELF structures by offset, and passes only bounded code-fragment snapshots to `objdump`. It never executes or loads target code and does not copy the full target. Detected input mutation makes the report incomplete: ```sh python3 scripts/brotli_binary_triage.py /absolute/target > triage.json ``` The default mode selects defined `Decode*BlockSwitch`, `SafeDecode*BlockSwitch`, and `ProcessCommands*` symbols, with `BrotliDecoderDecompressStream` as a public-symbol fallback. It records the target SHA-256, exact commands, selected scope, limits, symbols, and bounded output. Fragment hashes and source offsets bind the assembly to selected bytes; raw objdump output uses `.data`, not original section names. Resolve numeric call targets using the reported symbol addresses and original ELF scope. Exit 0 means evidence collection completed, not that the binary is fixed. Exit 3 means the evidence is incomplete. For stripped code, use only a range derived from trusted link/map or scanner evidence, for example `--file-offset OFFSET --length BYTES`, or `--start-address START --stop-address STOP [--section .text]`. Do not guess a range from a string location. Symbol and explicit-range modes both enforce `--max-range-bytes`; raise that bound deliberately when a trusted symbol is larger. If a trusted tool or defensible code scope is unavailable, keep the result unknown and state that missing evidence and the exact next request. Review triage evidence against the reviewed affected and fixed `decode.c` sources, not against mnemonic or string presence. Trace each literal, command, and distance block-switch path. The fixed design returns `BrotliDecoderErrorCode`, produces `BROTLI_DECODER_ERROR_FORMAT_BLOCK_SWITCH` for the invalid switch, and propagates that status through `ProcessCommands*`; the old design uses boolean/void handling that can lose the specific failure. Confirm the safe and non-safe callers and the error propagation, not merely a helper body. A version constant, import/export, function name, or nearby error string cannot establish the fix. If optimization, inlining, missing callers, output limits, multiple copies, or unfamiliar control flow prevents that trace, preserve `unknown`. One fixed-looking range never clears other copies or establishes runtime use. For artifact hashes and rollout inventory, optionally use: ```sh python3 scripts/brotli_binary_inventory.py /absolute/target ``` Its exit codes differ: 1 means inventory candidates (including fixed matches), 3 means coverage gaps, and 0 means no candidates/gaps, not that an app is safe. End-of-life Node.js versions get supported-branch upgrade advice, while decoder patch status remains unknown (exit 3) without build or vendor evidence. Validation has covered decoder source references, synthetic ELF/packaging fixtures, an opt-in suite for locally compiled stripped ELF artifacts, and a benign loopback test of built-in Node fetch response decompression. No deployed application or production binary's patch state was validated, and no malformed or exploit payload was executed. The benign fetch test establishes a decoding route, not CVE triggerability or exploitation. Do not describe source classification as an end-to-end test of a vulnerable app. The packaged function-section profiles were additionally validated with real linked `--gc-sections` builds for x86-64 GCC 14.2, AArch64 GCC 14.2, and AArch64 GCC 16.2 across all six releases and four optimization/PIC profiles. LoongArch GCC 14.2 complete references received the same 24-case linked/QEMU matrix. These benign executions confirm ordinary decoder operation and reference recognition; they are not malformed-stream, reachability, or vulnerability tests. ## Establish patch status and reachability Read [the advisory and detection reference](references/advisory.md) for source provenance, fixed branches, backports, and detection limits. Verify current vendor guidance if online research is available; otherwise date the assessment to the packaged snapshot and flag that advisory freshness is unverified. Inspect lockfiles, submodules, downloaded build inputs, native libraries inside archives, linking configuration, SBOMs, and the shipped binary. Check all build variants relevant to the deployment. Do not equate a dependency declaration or an inventory entry with compilation or shipment. Stripped/static binaries need build provenance, vendor evidence, or binary analysis; absence from a dynamic library listing is insufficient. A fixed OS library does not fix embedded copies. When relevant source is available, use Trailmark to accelerate discovery of decoder/fetch calls and their callers. From this skill's directory, run the packaged helper with uv, which installs its pinned Trailmark 0.7.0 dependency: ```sh uv run --script scripts/brotli_source_discovery.py /absolute/source > discovery.json ``` Choose the relevant source subtree; initial parsing is not resource-bounded. The first use may download Trailmark and its parser grammars. Do not substitute an unverified ambient `trailmark` command or run tools supplied by the target. The helper reads source without importing or executing the target. Keep the inventory scan: Trailmark skips dependencies, build outputs, and hidden folders, and can omit top-level JavaScript/TypeScript calls without a diagnostic. Use `candidates[].call` and the `neighborhood` nodes, calls, and entrypoint tags to prioritize source inspection. Preserve callsite locations, unresolved targets, confidence, deferred callback edges, and `configured_by` provenance. Verify bindings and callers in source; static confidence does not establish execution. For a reviewed import alias or wrapper, rerun with `--symbol NAME`. This adds a search target, not a verified binding. Repository `.trailmark` links and entrypoint overrides are declarations that require independent review. Exit 0 means evidence was collected, not that a decoder is fixed or reachable. Exit 3 means collection is incomplete; inspect diagnostics, truncation, and gaps. Missing paths or entrypoint tags never establish unreachability. If uv, parsers, or supported source are unavailable, state the gap and continue manual review. See [source discovery](references/reachability.md#trailmark-source-discovery) for scope and review steps. For each relevant surface, use [the reachability guide](references/reachability.md). Trace the actual input entry point through middleware/wrappers to the selected native decoder. Record configuration, authentication before decompression, input ownership, and process boundaries. Consider inbound requests, outbound responses, uploads and WOFF2/fonts, Java JNI stacks, workerd, and negotiated TLS certificate compression. Pure Java/Rust implementations require their own evidence; a language or algorithm name alone does not establish this C bug. For Node.js apps, explicitly investigate built-in `fetch()`, `globalThis.fetch`, and Undici wrappers, including attacker-controlled URLs and redirects. Automatic Brotli response decoding can reach the native decoder without an explicit Brotli dependency or `node:zlib` import. Establish the actual runtime's patch state. The offline inventory records source-reference and dependency leads. Use Trailmark's call evidence to find wrappers and callers, then verify fetch bindings, URL ownership, and the running Node version yourself. Review source locations and continue with the Node/fetch section of the reachability guide. A fixed supplied version or standalone Brotli library does not clear an unresolved runtime lead. For non-C/C++ projects, use the [ecosystem integration guide](references/ecosystems.md). The inventory's `decoder_source_candidate` findings collect language-scoped imports and decoder API references with source locations. Their `implementation_hint` is a lead based on expected upstream behavior, not proof of the selected provider. Resolve aliases, optional dependencies, runtime configuration, native extensions/JNI/FFI, and WASM assets before patch assessment. Inspect bundled native code and deployment-specific shared providers; do not substitute the review host's libraries. Independent Java, Rust, Go, and JavaScript ports require their own implementation review, not the C library's version rules. The registry includes separately reviewed exact Go/Rust source fingerprints for analogous block-switch error fixes. Preserve their `issue_relation` and upstream fix evidence; do not describe those matches as proof of the C corruption impact. Keep wrapper version, decoder version, build selection, shipment, and input ownership separate. Use reviewed local aliases via source discovery's `--symbol` when its default names cannot follow a call. Absence of a lexical lead or graph edge is not evidence that decoding cannot occur. For Electron apps, use the guide's Electron section. Separate Chromium's renderer/net/session fetch paths from Node fetch, and verify which decoder the embedder links. A standalone Node version rule cannot classify a vendor Electron build. Mach-O and ASAR packages require additional inspection. Benign Electron and Slack renderer measurements confirmed local response decoding, while Slack's shipped decoder patch status and attacker-controlled product paths remain unknown. The offline scanner does not launch either application. Report reachability as `confirmed`, `not_reachable_in_reviewed_configuration`, or `unknown`, with paths/lines or deployment evidence. A negative conclusion must name the exact configuration and reviewed paths. Missing evidence remains unknown. Do not infer exploitability from reachability, CVSS, a crash, or a third-party PoC. ## Explain the result and next action Lead with a short answer that matches the evidence: - "Update needed" when the app is confirmed to use an affected copy. State separately whether outside input can reach it. - "Vulnerable Brotli source found" when affected code is present but its use by the running app is unknown. Say "We still need to check whether your app uses this copy." Do not label the whole app vulnerable from source presence alone. - "Fix confirmed for the checked copy" when supported by verified evidence. State the scope; other copies and the running app may still need checking. - "More information needed" when the result is inconclusive. Explain the specific missing information and who can obtain it. Follow with what to do next. Name the affected component and the applicable update when established, who should make it, and how to confirm the running app has it. For a bundled copy, explain that the app must be rebuilt and deployed again; updating the computer's shared library may leave that copy unchanged. If action needs a maintainer or hosting provider, provide a short, ready-to-send message with the specific finding and requested update/confirmation. Do not send it yourself. For several copies, use a small table: location, plain-language result, next action. A fixed copy must not hide a vulnerable copy. Prioritize affected shipped copies with confirmed untrusted input, then resolve other candidates. State important unchecked files or missing running-app evidence in ordinary language. An empty scan is inconclusive. Do not label it "safe" or "unaffected". Keep the initial answer to a few sentences plus any useful short table. Avoid raw status codes, hashes, long evidence columns, and unexplained terms such as "shipment", "reachability", "JNI", or "provenance" in that answer. Preserve the full per-copy evidence (patch confidence, build selection, shipped artifact, input path, paths/lines, advisory date, and source/vendor links) for a maintainer or a request for technical details. Do not make the user interpret a security audit table to decide what to do. Where an update cannot be deployed immediately, identify the specific receiving decoder path that could be disabled and verify that boundary. Response encoding settings alone do not establish that request, font, or certificate decoding is disabled. Never present the community source inventory as vulnerable deployments.
© trailofbits · MIT · trailofbits/brotli-checker