Trust is a first-class feature of Uniqent, not an afterthought. This document covers both how the toolchain protects users and how to report vulnerabilities.

• Secret-scan gate (fail-closed). pack, validate, and sign run scanForSecrets() over every file in a bundle, using entropy heuristics plus known key prefixes (sk-, ghp_, xoxb-, …). Any likely secret value fails the operation. Secrets are never permitted in a bundle — only credential references and install-time resolution.
• Signing & verification. Bundles are signed with Ed25519 over a canonical digest of their contents. verify fails if any file changed after signing. Installing an unsigned or unverifiable bundle requires an explicit --allow-unsigned flag and prints a loud warning.
• Permission sheet + memory preview. Before any write, install shows the requested filesystem, network, autonomy, and process-spawn scopes, and lets the user redact or skip memory items.
• Sandboxed dry-run. Install loads the agent in a sandbox and confirms tools/MCP/persona/memory before apply() writes anything to the real framework.
• Skill script static scan. validate statically flags skill scripts that shell out, make network calls, or look obfuscated (warning in v1; blocking on registry publish later).

Please report suspected vulnerabilities privately rather than opening a public issue. Email the maintainers (see repository metadata) with a description, reproduction steps, and impact. We aim to acknowledge within a few business days and will coordinate a fix and disclosure timeline with you.

Do not include real secrets or credentials in reports or test bundles.