convez.ai
AgenteSegurançaAvançado
websentry-ai

OACB — Framework de Segurança para Agents Autônomos Claude

websentry-ai/oacb

Hooks de segurança para Claude Code autônomo com auditoria e bloqueio em camada host

Ver no GitHub
BomPluginScore · de 0 a 100

Framework de segurança que instala hooks e controles em modo autônomo do Claude Code, bloqueando comandos destrutivos, injeção de prompts e violações de allowlist. Oferece quatro níveis de enforcement (shadow, baseline, strict, paranoid) mapeados aos padrões OWASP ASI 2026 e MITRE ATLAS.

Instalação

$npx convez-ai websentry-ai/oacb

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
62%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Proteger ambientes autônomos contra comandos destrutivos (rm -rf, terraform destroy)
  • Auditoria e conformidade de ferramentas MCP em modo auto
  • Rollout gradual com modo shadow para observação antes de enforcement

Veredito

Framework robusto de segurança em camada host para agents autônomos, com tiers flexíveis e auditoria. Reconhece honestamente suas limitações (obfuscação, egresso de rede) e é adequado para ambientes regulated, mas ainda em beta — requer observação em shadow antes de produção.

README

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

OACB — Open Autonomous Coding-agent Baseline

Security hooks and settings for Claude Code running in autonomous (auto) mode. Enforces strict controls to prevent accidental or malicious damage by autonomous agents.

License: Apache 2.0 OACB v0.2.1 beta Status: Beta

Beta release. OACB is under active development. Interfaces — rule IDs, audit-log JSON schema, managed-settings keys, tier semantics — may change before v1.0.0. Recommended for evaluation, shadow-tier observation, and pre-production rollout. Pin a specific tag (v0.2.1) in fleet deployments; do not track main in production.


Quick start

npm install -g unbound-cli

unbound oacb check                    # verify your Claude Code install
unbound oacb apply                    # shadow tier - log-only, safe first step
unbound oacb doctor                   # confirm hooks are working

After a week or two in shadow mode, graduate to enforcement:

unbound oacb apply --tier baseline    # block destructive commands

What OACB does

OACB installs shell hooks and a settings overlay into ~/.claude/ that:

  • Block destructive Bash commands (rm -rf, terraform destroy, force-push, DB wipe) before they run
  • Block Claude Code's --dangerously-skip-permissions bypass at all tiers
  • Warn or block prompt injection patterns in user input (bidi Unicode, instruction-override markers)
  • Audit every MCP tool call; block allowlist violations at strict/paranoid
  • Log all decisions to ~/.claude/hooks/oacb-audit.log

Tiers

TierWhat it doesWhen to use
shadowLog only — never blocksObserve for 1–2 weeks before rollout
baselineBlock CVE-mapped destructivesDefault for most teams
strictBlock all; human-gate on ambiguousRegulated environments, pre-prod
paranoidAuto mode disabled; managed hooks onlyFedRAMP, high-sensitivity
unbound oacb apply --tier shadow
unbound oacb apply --tier baseline
unbound oacb apply --tier strict
unbound oacb apply --tier paranoid

Other commands

unbound oacb audit                    # score your current settings, report gaps
unbound oacb diff --to baseline       # preview what changes at the next tier
unbound oacb doctor --verbose         # run the conformance suite against installed hooks
unbound oacb remove                   # surgically remove OACB from settings.json

Every tier's rule set is open JSON — pass --overrides <path> to apply to layer your own deny rules, allowlist additions, or MCP policies on top of the baseline. This makes OACB adaptable to your organisation's specific tooling and risk posture without forking the project.


Honest limits

OACB is a host-layer control. Pattern-match rules can be defeated by obfuscation the agent controls (\rm, base64-pipe, MCP indirection, heredocs). The hooks raise the bar considerably, but they do not make Claude Code airtight.

The gap that host-layer controls cannot close is network egress and LLM traffic. For that you need enforcement below the agent — inline on the model API call — where every tool invocation, prompt, and response can be inspected and blocked before it reaches the filesystem or network.

Unbound Gateway is an Agent Access Security Broker (AASB) built for exactly this: inline policy on LLM traffic, egress DLP, prompt-injection detection, and cross-agent audit — all without changing your Claude Code workflow. OACB + Unbound Gateway is the full-stack answer: OACB handles the host, Gateway handles the network and LLM layer.

Known host-layer residuals are documented in non-claims.md.


Contributing

Security reports: security@unboundsecurity.ai — 72h ack, 90-day coordinated disclosure. See SECURITY.md.

Code contributions: open a PR. New rules require (a) an OWASP ASI mapping, (b) a passing adversarial test case, (c) false-positive coverage.


Acknowledgements

Rules are grounded in published research by GMO Flatt Security (CVE-2025-66032), Check Point Research (CVE-2025-59536), Cymulate (CVE-2025-54794/5), Embrace The Red (CVE-2025-55284), Invariant Labs, Adversa AI, Pillar Security, Trail of Bits, and others. If you publish a Claude Code bypass, we will cite you and ship the test case.


OACB is maintained by Unbound Security and released under Apache 2.0.

Ver o original no GitHub

Bom para

Relacionados