Security

Threat model

Runtime Guard is designed to guard against unsafe behavior from AI agents operating on your system. The primary threat categories we address:

Prompt injection — Crafted inputs that attempt to override agent instructions or bypass safety controls.
Tool abuse — Agents using shell, filesystem, or network tools outside their defined scope or policy.
Data exfiltration — Unauthorized outbound transfer of credentials, keys, or sensitive files to untrusted destinations.
Privilege escalation — Attempts to gain elevated access, modify system configurations, or install persistence mechanisms.
Unsafe plugins — Third-party extensions or plugins that violate access policies or behave unexpectedly.
Remote code execution — Piped scripts, inline code evaluation, or downloads from untrusted sources.

Shell commands — Every command executed by an agent is intercepted and evaluated against your policy.
Filesystem access (reads/writes) — File operations are tracked to detect unauthorized access to sensitive paths.
Outbound network requests — Network calls are inspected for data exfiltration and connections to untrusted destinations.

Policy-driven blocking examples:

SSH key access blocked — Agents cannot read or copy your SSH keys or credentials.
Untrusted outbound blocked — Network requests to unapproved domains are denied.
Startup folder writes blocked — Agents cannot install persistence mechanisms or modify startup configurations.
Privilege escalation blocked — Attempts to elevate permissions or modify system-level configurations are prevented.

We believe in transparency. Here is what Runtime Guard does not do:

Planned Behavioral baseline detection — Learn normal agent behavior and flag deviations.
Planned Anomaly scoring — Score actions based on deviation from established baselines.
Planned Signed policy bundles — Tamper-resistant, portable policy packages.

Questions about our security approach? Contact us at security@airuntimeguard.com