Agent Guard

Giám sát tác nhân AI, bảo vệ dữ liệu nhạy cảm và yêu cầu phê duyệt thao tác rủi ro cao.

Agent Guard helps you supervise, understand, and control high-risk actions performed by AI agents on websites. On sites you explicitly authorize, Agent Guard can observe supported browser-visible actions, identify sensitive-data categories, detect prompt-injection risk signals, compare actions with your declared intent and policies, and require approval before supported high-risk operations proceed. KEY FEATURES • Live agent supervision View supported browser-visible actions such as navigation, clicks, form input, submission attempts, uploads, downloads, authorization, publishing, account changes, purchases, payments, and deletion attempts. • High-risk action approval Require approval before supported sensitive or irreversible actions proceed. Review the action, destination, sensitive-data categories, reversibility, policy result, intent mismatch, and available protection coverage. • Intent Contracts Define what an agent is allowed to do during a session, including permitted sites, actions, destinations, sensitive-data categories, spending limits, upload limits, recipient limits, and other risk budgets. • Guard Profiles Choose Observe, Balanced, Strict, or Custom protection modes. • Sensitive-data protection Detect categories such as names, email addresses, phone numbers, addresses, authentication information, payment information, identity information, health information, location, API tokens, private keys, and uploaded files. Agent Guard is designed to retain redacted categories and minimized evidence rather than raw sensitive values. • Prompt-injection risk signals Identify signals such as hidden text, visually concealed content, command-like instructions, requests to reveal credentials, instructions to ignore prior directions, and suspicious cross-origin instructions. Prompt-injection findings are heuristic risk signals. Agent Guard does not claim to detect every attack or prove that a detected instruction caused a later action. • Guard Health See which channels are guarded, observed only, unsupported, or unavailable. Agent Guard clearly distinguishes actions it can pause from actions it can only observe. • Policy Center and Policy Lab Create rules by site, action, destination, data category, and agent type. Test proposed rules against redacted historical events before applying them. • Data-flow and evidence views Review redacted sensitive-data flows, new destinations, prompt-signal associations, policy decisions, approval history, and differences from earlier sessions. • Panic Lock Immediately cancel pending approvals, expire temporary grants, and block supported high-risk actions in the current session. • Optional cross-tab workflow coverage With optional permission, associate related OAuth, payment, authorization, and newly opened tabs with the current guarded workflow. • Optional Network Quarantine With optional permission, create temporary, tab-scoped rules that block specifically identified third-party destinations during a guarded session. • Local audit and export Keep bounded audit information on your device and export redacted reports as JSON, CSV, HTML, or an incident bundle. PRIVACY BY DESIGN Agent Guard is local-first. • No advertising or behavioral profiling • No sale of user data • No developer-operated analytics or telemetry • No upload of audit records or browsing history to the developer • No remotely hosted executable code • Site access is requested at runtime for origins selected by the user • Sensitive values are minimized and redacted at collection, storage, and export boundaries OPTIONAL PERMISSIONS Optional features may request additional permissions only when enabled: • Notifications: generic notices for pending high-risk approvals • Web navigation: advanced cross-tab workflow and document-lifecycle coverage • Network quarantine: temporary, scoped blocking of identified destinations LIMITATIONS Agent Guard supervises supported browser-visible behavior. It cannot observe every AI agent, every browser or server-side action, encrypted downstream processing, server-to-server transfers, or every prompt-injection technique. Some channels may be observed but cannot be reliably paused. Website changes, browser restrictions, cross-origin frames, service workers, cached responses, WebSockets, sendBeacon, and server-side processing may limit coverage. Agent Guard provides evidence, deterministic policy decisions, and heuristic risk signals. It is not an antivirus product, a password manager, a legal or compliance certification, or a complete data-loss-prevention system. High-risk actions should still be reviewed carefully before approval.