convez.ai
MCP ServerSegurançaAvançado
Rul1an

Plimsoll - Revisão de Lançamento de Agentes IA

Rul1an/plimsoll-action

Diff de capacidades de agentes IA + gate de segurança para releases

Ver no GitHub
BomPluginScore · de 0 a 100

MCP server que analisa mudanças de runtime em lançamentos de agentes IA (filesystem, rede, ferramentas MCP) entre versões, gera diffs de capacidade e bloqueia PRs até aprovação humana. Integrado ao GitHub Actions com relatórios em code scanning e proteção contra tool poisoning e rug pulls.

Instalação

$npx convez-ai Rul1an/plimsoll-action

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
71%
Último commit
há 2 semanas
Versão
v1.1.6
Issues abertas
6
Contribuidores
2

Casos de uso

  • Revisar mudanças de runtime em agentes antes do deploy
  • Detectar tool poisoning e comportamentos não autorizados
  • Gate de segurança em CI/CD para equipes de plataforma IA
  • Relatórios de supply chain em code scanning

Veredito

Ferramenta especializada e rigorosa para segurança de agentes IA em produção, com abordagem fail-closed e observação de efeito real (não apenas descrições). Essencial para times que shipar agentes MCP com requisitos de compliance e prevenção de supply chain attacks.

README

A tradução para o português deste README ainda está na fila.

Plimsoll release review (GitHub Action)

Marketplace

See what an AI agent release changed at runtime, review it, and gate the PR. On every pull request, Plimsoll diffs what your agent can actually do (filesystem, network, MCP tools) between the previous release and the new one, decides whether a human should approve the change, posts a review on the PR, and reports findings in code scanning.

Built for AI platform teams and security engineers shipping agents (MCP servers, tool-using agents, agent workflows) who need to answer one question before deploy: what runtime capability changed, and should we approve it?

Topics: AI agents, agent security, MCP, tool poisoning, rug pull, runtime capability, release review, capability diff, supply chain, SARIF, code scanning, CI gate.

Quick start

# .github/workflows/plimsoll.yml
name: Plimsoll release review
on: pull_request
permissions:
  contents: read
  pull-requests: write
  security-events: write
jobs:
  review:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4         # pin to a SHA in production
      - uses: Rul1an/plimsoll-action@v1   # pin to a SHA in production
        with:
          before: before.json            # previous release's capability surface
          after: after.json              # the new release's capability surface
          fail-on: pending               # fail until a human approves new capability

A capability surface is what the agent actually did at runtime. It comes from assay (kernel-observed via eBPF) or any producer of the simple {fs, net, tools, procs} shape. The action also accepts an assay runner archive (.tar.gz).

What you get, free, no account

Runs on any repo with no token, from the open review engine in this repo:

  • a capability diff of the two releases (new filesystem paths, network endpoints, MCP tools),
  • a sticky review comment on the PR and a job summary,
  • SARIF uploaded to code scanning, so findings show in the Security tab and PR annotations,
  • a configurable gate: fail-on: pending holds the release until a human approves new capability; blocked fails only on a coverage hold; never reports without failing,
  • coverage honesty, fail-closed: a no-diff run is only auto-cleared for surfaces actually observed. If a relevant surface (say the network layer) was not observed, the release is held as inconclusive_observation_gap, not passed quietly. "Saw nothing" is never read as "did not look".
  • evidence hygiene: if a recorded surface value looks like a secret (provider tokens, PEM private keys, JWTs, bearer tokens, key=value credentials), the review warns and emits a PLIMSOLL-POSSIBLE-SECRET finding so you can redact it at capture. It is value-free (it never echoes the matched value) and advisory, not a gate.

This is the whole review for one repo, not a teaser.

Catching a tool that changed after you approved it

The catch comes from observed effect, not from re-reading a tool's description, so it holds up against tool poisoning and rug pulls (a tool whose description or behaviour changes after approval). A translation agent that quietly starts reading a credentials file, opening an undeclared endpoint, and using a new outbound-HTTP tool is surfaced as three findings and held as pending, while a routine release that only touched its workspace auto-clears.

Maximum use (the hosted layer)

The free action reviews one repo at a time. Teams running many agents can connect a run to the hosted layer with the license input: a cross-repo audit ledger with retention, a fleet view of capability drift, a governed approval workflow (RBAC, SSO, approval queues), and anchored review attestations. Those need a hosted service, so they live behind a Plimsoll account rather than a crippled local check. The open review engine stays open. The hosted layer is in preview.

How this differs from the assay action

assay actionPlimsoll action (this one)
Inputevidence bundles from agent runstwo capability surfaces (previous vs new release)
Doesverify, lint, diff vs baseline, compliance packs, attestcapability diff, review policy, approve/hold, gate
Answersis this run's evidence valid and complete?what changed between releases, and do we approve it?

They compose: assay verifies the record is trustworthy; Plimsoll reviews the capability change and gates it. assay is observe-and-verify; Plimsoll is review-and-decide.

Security

  • Pin this action and every action in your workflow to a full commit SHA, not a tag.
  • No third-party action dependencies: it uses python3 and the gh CLI on the runner and talks to the GitHub API with the job token only.
  • On public repos code scanning is free; on private repos the SARIF upload needs GitHub Advanced Security, and is skipped gracefully without it (the gate and comment still run).

Inputs and outputs

See docs/GITHUB-ACTION.md for the full reference and examples/plimsoll-review.yml.

License and source

Apache-2.0. This repo holds the open review engine plus the action. The broader Plimsoll product lives elsewhere; the review primitives are the open part.

Ver o original no GitHub

Bom para

Relacionados