Habilidades mantis-meta-agent

📦

mantis-meta-agent

Name: mantis-meta-agent
Author: google

Revisão do conteúdo r1 Crítico ⚙️ Comandos externos📁 Acesso ao sistema de arquivos

Coordinate Continuous Mantis Security Reviews

Long security reviews require coordinated specialist stages and reliable campaign state across repeated passes. This skill supervises Mantis subagents, preserves findings, and manages optional synchronization and snapshot provenance.

Suporta: Claude Codex Code(CC)

⚠️ 38 Ruim

Não instale este Skill automaticamente.

A política canônica exige a revisão de um operador antes de qualquer ação de instalação.

Instalar com meu Agente

Copie esta solicitação para o seu Agente. Ela inclui a página canônica do Skill e o manifesto.

Solicitação do agente

Review the Skillstore skill "mantis-meta-agent" from https://skillstore.io/skills/google-mantis-meta-agent.md and its manifest at https://skillstore.io/api/skills/google-mantis-meta-agent/manifest. Verify the artifact. Do not auto-install. Inspect the skill and report your findings, then wait for an operator or manual installation decision.

Seu Agente ainda deve apresentar o plano e solicitar qualquer confirmação exigida pela política de segurança.

Recursos legíveis por agentes

Use estes links quando um agente de IA, crawler ou script precisar de contexto limpo em vez de ler a página completa.

Detalhe em Markdown GET /skills/google-mantis-meta-agent.md manifest assinado GET /api/skills/google-mantis-meta-agent/manifest lockfile assinado GET /api/skills/google-mantis-meta-agent/lockfile

Testar

Usando "mantis-meta-agent". What should happen when an opt-in sync finds local repository changes?

Resultado esperado:

Illustrative status: synchronization skipped because local changes exist.
Preserve the current tree and report why synchronization did not run.
Attempt the requested snapshot only under the campaign's pinning rules.

Usando "mantis-meta-agent". What should a completed pass preserve for the next review?

Resultado esperado:

Expected archive contents: finding records, campaign state, and a copied knowledge base.
Expected live state: the knowledge base remains available and the pass counter advances.
Expected report: reproduced actionable findings, supporting evidence, and patch verification results.

Usando "mantis-meta-agent". The harness exposes no subagent execution logs. What should reflection receive?

Resultado esperado:

Expected delegation: provide an explicit empty log list and request a missing-transcript insight for each executed stage.

Auditoria de segurança

Crítico

v1 • 04/10/2026 Abrir o relatório versionado

Most static findings are Markdown formatting, campaign state references, or legitimate repository checks, not shell execution or system reconnaissance. Confirmed concerns include unproven backup ownership and unredacted transcript access; semantic findings identify self-declared testing authorization and permissive snapshot verification. No evidence found of credential exfiltration, malware installation, or instructions to manipulate this marketplace audit.

Arquivos analisados

612

Linhas analisadas

Itens de revisão

Falsos positivos ignorados

Preocupações de segurança confirmadas (2)

Crítico

SKILL.md:111-120 SKILL.md:469-481

Prompt Injection Attempt Detected

The directive states, "You are authorized and encouraged to use whatever suitable tools are at your disposal," including testing live endpoints. This self-declared authorization can mislead agents into testing outside user-approved scope or running untrusted targets without isolation.

The text explicitly grants broad tool-use authorization and includes live endpoints, reproduction, and exploit-chain stages without defining permission or isolation boundaries. Defensive intent does not establish authorization.

Médio

SKILL.md:293-303 SKILL.md:326-341

Snapshot Verification Accepts Incomplete or Writable Targets

Snapshot verification permits pinning when missing files are not already referenced and labels write protection as best effort. Downstream stages can treat an incomplete or still-writable target as pinned evidence.

The instructions explicitly permit skipped files and best-effort chmod before setting pinned status. This contradicts the claimed immutable, authoritative snapshot guarantee.

Itens de revisão de capacidade (6)

Estas são capacidades locais reais que podem ser esperadas para esta habilidade, portanto exigem revisão, mas não são contabilizadas como comportamento malicioso confirmado.

Médio

SKILL.md:173

Ruby/shell backtick execution

`find <target_root> -regextype posix-extended -regex '.*\.bak-[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0

The find command deletes UUID-shaped backups without requiring proof that Mantis created them. User-owned backups can match despite lock and state-directory exclusions.

Médio

SKILL.md:182

Ruby/shell backtick execution

invisible to every dirty check. Delete stray `mantis-summary.md` and

Pre-sync cleanup deletes mantis-summary.md by filename before checking repository changes. A user-owned or modified file with that name can be lost.

Médio

SKILL.md:183

Ruby/shell backtick execution

UUID-shaped `*.bak-<uuid>` files at the LIVE root first (same narrow

Cleanup deletes UUID-shaped backups at the live root based on their names. The instructions do not require provenance proving these files are disposable.

Médio

SKILL.md:184

Ruby/shell backtick execution

glob as the SCOPE rule above — never the broad `*.bak-*` glob for

The narrow glob reduces accidental matches but still permits deleting user-owned UUID-shaped backups. Filename shape alone does not establish Mantis ownership.

Médio

SKILL.md:173

Hidden file access

`find <target_root> -regextype posix-extended -regex '.*\.bak-[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0

The find command deletes UUID-shaped backups without requiring proof that Mantis created them. User-owned backups can match despite lock and state-directory exclusions.

Médio

SKILL.md:498

Hidden file access

`<appDataDir>/brain/<conversation_id>/.system_generated/logs/transcript.jsonl`,

The reflection stage receives full campaign execution transcripts from harness-private storage without mandatory redaction. Logs can contain credentials or private context beyond needed learning summaries.

Fatores de risco

⚙️ Comandos externos (50)

📁 Acesso ao sistema de arquivos (27)

Auditado por: codex

Compartilhar e citar este relatório

Compartilhe o relatório de avaliação versionado, o badge neutro, o cartão incorporável e as citações. A Skillstore apresenta evidências sem decidir se este Skill é seguro.

Abrir o relatório versionado

Copiar link do relatório

https://skillstore.io/skills/google-mantis-meta-agent/audits/1?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Badge Markdown

[![Skillstore security assessment](https://skillstore.io/badges/skills/google-mantis-meta-agent/security.svg)](https://skillstore.io/skills/google-mantis-meta-agent?utm_source=security_passport_badge)

Badge HTML

<a href="https://skillstore.io/skills/google-mantis-meta-agent?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/google-mantis-meta-agent/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

Cartão incorporável

<iframe src="https://skillstore.io/embed/skills/google-mantis-meta-agent.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

Citações acadêmicas (APA · BibTeX · CFF)

Citação APA

google. (2026). mantis-meta-agent security audit report (audit version 1) [Author version unspecified]. Skillstore. https://skillstore.io/skills/google-mantis-meta-agent/audits/1

Citação BibTeX

@techreport{google-google-mantis-meta-agent-2026,
  author = {google},
  title = {mantis-meta-agent security audit report (audit version 1)},
  institution = {Skillstore},
  year = {2026},
  number = {1},
  url = {https://skillstore.io/skills/google-mantis-meta-agent/audits/1},
  note = {Author version unspecified}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "mantis-meta-agent security audit report (audit version 1)"
version: "unspecified"
type: report
authors:
  - name: "google"
date-released: "2026-10-04"
url: "https://skillstore.io/skills/google-mantis-meta-agent/audits/1"
identifiers:
  - type: other
    value: "skillstore:google-mantis-meta-agent:audit:1"
    description: "Skillstore immutable audit report identifier"

Pontuação do Skillstore

Por que esta pontuação Confiança das evidências: Baixo

Arquitetura

Manutenibilidade

Conteúdo

Comunidade

Conformidade com a especificação

O que você pode criar

Coordinate a Repository Review

Supervise specialist review stages and preserve validated findings across repeated passes on an explicitly authorized repository.

Track Review Provenance

Use optional synchronization, pass snapshots, and archives to compare campaign results across repository revisions.

Supervise an Artifact Campaign

Coordinate appropriate analysis specialists for an authorized binary or firmware artifact in an isolated testing environment.

Teste estes prompts

Inspect Campaign Readiness

Inspect readiness for a Mantis campaign on [authorized repository]. List missing subagents, tools, and state inputs. Do not start testing or modify files.

Supervise a Bounded Review

Supervise one Mantis pass on [authorized repository] using [external state directory]. Do not sync. Stop after archiving and summarize reproduced findings.

Request a Synchronized Pass

Run one authorized Mantis pass with --sync and --state_root=[external directory]. Preserve local changes and unknown artifacts. Halt if snapshot completeness or protection fails.

Resume and Steer a Campaign

Resume [authorized campaign] from its existing state. Check snapshot provenance and archive integrity. Focus the next pass on [component]. Require consent before sensitive-log processing.

Boas práticas

Define authorized targets, permitted testing methods, and stopping conditions before launching a campaign.
Keep campaign state outside the target tree, preserve unknown artifacts, and verify snapshot completeness before trusting provenance.
Use isolated reproduction environments and redact sensitive campaign transcripts before sharing them with reflection subagents.

Evitar

Treating the skill's authorization statement as permission to test arbitrary endpoints or execute untrusted binaries.
Deleting files solely because their names resemble generated Mantis artifacts.
Treating unpinned, incomplete, or writable snapshots as immutable evidence for authoritative findings.

Perguntas frequentes

Which AI tools are listed as supported?

The metadata lists Claude, Codex, and Claude Code. Actual execution requires a harness with compatible Mantis subagents and delegation tools.

Does this package include the complete review pipeline?

No. The reviewed package contains one instruction file. Specialist subagents, schemas, and supporting tools must be available separately.

Does the supervisor patch vulnerabilities directly?

No. It delegates patching and reproduction to specialist subagents and coordinates their results.

Does synchronization run automatically?

No. Synchronization and pinning require --sync or an explicit request. Default passes operate against the live target.

Can a campaign resume after interruption?

The instructions define pass recovery and snapshot reuse checks. Missing previously pinned snapshots should stop the campaign rather than silently replace evidence.

What safeguards are needed before use?

Confirm testing authorization, isolate untrusted execution, verify artifact ownership before deletion, redact transcripts, and require complete protected snapshots.

Detalhes do desenvolvedor

Autor

google

Licença

MIT

Revisão da Skillstore

Aviso de versão

O autor não declarou uma versão.

Repositório

https://github.com/google/mantis/tree/f48a85e8823ee6f4824ae9a24b9b2efb8aab8c9a/mantis-meta-agent

Referência

44cf00a7ced88385654599ad1ed145fec23ee8ed

Atualidade da manutenção

05/10/2026

Uso

0 downloads · 0 visualizações

Estrutura de arquivos

📄 SKILL.md