Company policy. Local action. Reviewable evidence.
Set rules for AI tools, agent autonomy, connections and sensitive data. Apply supported controls, assign findings and keep the version behind every decision.
Policy checks and live request controls work at different moments. Both should explain what was checked, what happened and why.
Define approved or disallowed tools, permitted agent modes and MCP hosts. Each publish creates a new policy version. Device reports are evaluated against the rules in force.
Allow, warn, redact or block supported AI interactions. Prompt checks can detect secrets and personally identifiable information before a request is sent.
Tie the result to the policy version that produced it. Assign findings, track remediation and retain evidence your security and governance teams can review.
These are different outcomes. Choose the appropriate rule and confirm support for the application and environment you intend to govern.
Grade its presence against company policy and surface a finding. A disallowed tool can remain installed.
Policy findingApply allow, warn, redact or block on supported interactions. 31 built-in detectors for personal data, secrets and keys, plus your own patterns, in Claude Code, Cursor, Codex CLI and a browser extension for Chromium browsers. Coverage depends on the tool and enforcement path.
Runtime controlA blocked-app workflow requests administrative approval before removal. Do not confuse it with blocking one prompt. Kill switches halt a tool, or the whole fleet, at once.
Approved remediationReview permitted modes and the Guardian plugin’s 13 controls, prompt injection and jailbreak among them, in Claude Code, Cursor, Codex CLI and Antigravity. Computer-use rules cover agents that drive a browser or the desktop, or send email and chat as the person: each is Off, Warn or Block.
Validate host supportSelect an outcome. These controls explain the behaviour; they do not change a real policy.
The configured rule permits this supported request. Keep the decision linked to the policy version.
Framework packages help organise rules and evidence for SOC 2, the EU AI Act, ISO/IEC 42001 and NIST AI RMF. They do not certify your organisation.
Keep the rule, its version and the resulting verdict together. A later policy change does not rewrite an earlier decision.
Separate pass, warn, fail, unknown, ungraded and not applicable. Missing evidence must never inflate a pass rate.
Use evidence packs, exports and retention settings to support your review process. Prompt inspection and stored conversation content are separate choices.
Pilot your target apps, operating systems and workflows. Confirm supported actions, permissions, compatibility and offline behaviour before expanding enforcement.
Plan an agent pilot →See the Guardian Agent Platform on your own fleet. Find every AI tool, understand what it reaches and costs, and apply the controls your team needs.