Skip to content

Security: ali-abassi/agent-workflows

SECURITY.md

Security policy

Supported version

Security fixes are applied to the latest commit on main.

Reporting a vulnerability

Do not open a public issue for a vulnerability. Use GitHub's private vulnerability reporting for this repository:

  1. Open the repository's Security tab.
  2. Choose Advisories.
  3. Choose Report a vulnerability.

Include the affected command or workflow, impact, reproduction steps, and any suggested mitigation. Please avoid accessing data or systems you do not own.

Execution boundary

Agent Workflows executes workflow commands with the permissions and environment of the invoking user. It is an orchestration and evidence layer, not a sandbox.

  • Treat third-party steps.yaml, prompts, retrieved text, and tool output as untrusted.
  • Run untrusted workflows inside a container or similarly isolated account.
  • Give agent nodes the smallest practical tool allowlist.
  • Keep credentials out of workflow files, prompts, command arguments, logs, artifacts, and committed run bundles.
  • Review effectful commands before execution.

tools: restricts what Pi exposes to a model. It does not restrict shell commands, filesystem access, child processes, network access, or inherited credentials at the operating-system level.

There aren't any published security advisories