site-docs/security.md

Rama main · réplica de lectura del repositorio original.

# Security Policy

## Reporting a Vulnerability

**Do not open a public GitHub issue for security vulnerabilities.**

Report via [GitHub Security Advisories](https://github.com/koda-ai-studio/agent-bom/security/advisories/new).

We aim to respond within 48 hours and provide a fix within 7 days for critical issues.

## Supported Versions

| Version | Supported |
|---------|-----------|
| Latest  | Yes       |
| < Latest | No — upgrade to the latest release |

## Security Design

agent-bom has three distinct security postures:

- **Scanner mode** (`agent-bom agents`, `agent-bom fs`, `agent-bom check`) is read-only. It reads config files and queries public APIs (OSV.dev, NVD, EPSS, CISA KEV).
- **MCP server mode** (`agent-bom mcp server`) is read-mostly. Scanner, posture, graph, and audit tools are read-only; Shield write actions require admin role, `shield:write` scope, and an audit reason.
- **Proxy mode** (`agent-bom proxy`) is an execution and enforcement surface. It intentionally launches or connects to the target MCP server so it can inspect, block, and audit tool traffic in real time.

Across all modes, agent-bom never stores credential values — only their names appear in output as `***REDACTED***`.

For the product-wide local, endpoint, cloud, API/UI, and proxy data boundary,
see [Data Access Boundaries and Operator Control](deployment/data-access-boundaries.md).

For incident response when LLM provider keys may have been exposed, use the
[LLM Key Exposure Drill](deployment/llm-key-exposure-drill.md).

For copied shell commands and assistant-generated installer prompts, use
[Assistant Command Safety](reference/assistant-command-safety.md).

## Release trust and dependency controls

For dependency controls, extras audit coverage, parser fuzzing, and release
trust posture, see:

- [Supply Chain and Trust](architecture/supply-chain-and-trust.md)