SECURITY.md

Rama main · réplica de lectura del repositorio original.

# Security Policy

`clawd` holds your Telegram credentials, LLM keys, and personal data, and it executes
tools on your behalf. Security reports take priority over feature work.

## Reporting a vulnerability

Report privately through GitHub: **Security → [Report a vulnerability](../../security/advisories/new)**.
Do not open a public issue for anything exploitable.

Include the affected version or commit, reproduction steps, and the impact you see.
You will get an acknowledgment within 7 days and a status update when the fix lands
or the report is declined, with reasoning either way.

## Scope

Reports in these areas matter most:

- Bypassing the Telegram allowlist (getting a conversation, or any action, without being allowlisted).
- Bypassing the approval fabric in direct mode: a forged or third-party callback that approves an
  action, or a consequential action that runs without the approval required for its resolved chat
  mode.
- Escaping the code-execution sandbox through clawd's integration (host paths, network, or secrets reachable from guest code without the documented opt-ins).
- Defeating the argument scanners, the direct-mode exfiltration approval gate, or the SSRF
  blocklist. Group mode skips the exfiltration approval round-trip by policy; its argument scanners
  and SSRF blocklist remain security boundaries.
- Secret exposure: plaintext at rest where encryption is promised, or secrets surviving log redaction.
- Prompt injection that crosses a policy enforced in code. Odd or unhelpful model output is not a vulnerability.

Vulnerabilities in upstream dependencies belong upstream; open a regular issue here
asking for the version bump.

## Supported versions

The latest release and `main`.