Post-authentication observability for AI agents. See what your agent does after it authenticates.
Built for the Authorized to Act: Auth0 for AI Agents hackathon.
At RSAC 2026, five major vendors shipped agent identity frameworks. None of them tracked what the agent did after authentication succeeded (VentureBeat, March 2026). The OWASP Top 10 for Agentic Applications catalogues the risks that emerge precisely in this post-authentication space: agents hijacking goals (ASI01), misusing legitimately granted tools (ASI02), and abusing inherited privileges (ASI03).
Agent Observatory makes post-authentication AI agent behavior observable, auditable, and controllable -- using Auth0's existing identity primitives.
| Feature | Usage | OWASP Mapping |
|---|---|---|
| Token Vault | RFC 8693 token exchange for Google Calendar, GitHub, Slack | ASI02, ASI03 |
| Token Vault Interrupts | Step-up authorization as circuit breaker for high-risk ops | ASI09 |
| FGA Pattern (in-memory demo) | Service-level + scope-level access control (ReBAC pattern) | ASI03 |
| Universal Login | User authentication via @auth0/nextjs-auth0 | ASI03 |
| Risk | Code | Our Mitigation |
|---|---|---|
| Agent Goal Hijack | ASI01 | Token Vault limits operations to defined scopes |
| Tool Misuse & Exploitation | ASI02 | Every tool call risk-classified and logged |
| Identity & Privilege Abuse | ASI03 | FGA service + scope authorization + least-privilege scopes + post-auth monitoring |
| Agentic Supply Chain | ASI04 | Tool definitions pinned at build time |
| Unexpected Code Execution | ASI05 | No code execution tools exposed |
| Memory & Context Poisoning | ASI06 | Not currently addressed (would require RAG integration) |
| Insecure Inter-Agent Comm | ASI07 | Single-agent architecture, no inter-agent calls |
| Cascading Failures | ASI08 | Error capture prevents silent cascading |
| Human-Agent Trust Exploitation | ASI09 | Interrupt-as-circuit-breaker for write ops |
| Rogue Agents | ASI10 | Real-time behavior monitoring via Observatory |
By logging token exchange events alongside tool execution events, we can answer: "Which tool calls consumed credentials from Service X in the last hour?"
OAuth scopes encode what an agent can do. By classifying scopes into risk tiers (read-only vs. write vs. administrative), we compute a per-operation risk score. calendar.freebusy (read availability) carries lower risk than chat:write (send messages).
Auth0's TokenVaultInterrupt mechanism generalizes to any authorization boundary. When a risk threshold is exceeded, throwing an interrupt pauses agent execution and surfaces the decision to the user -- converting the post-authentication gap from a silent failure mode into an explicit control point.
The #1 developer pain point: Token Vault setup is a 10-step process with a single uninformative error (Federated connection Refresh Token not found). Our debugger shows:
- Per-connection token state and health scores
- Exchange timeline and error history
- Configuration checklist per service
- Known error reference (auth0-ai-samples#66, auth0-ai-js#175)
Every getAccessTokenFromTokenVault() call is instrumented to record: which tool requested the credential, what scopes were used, when the token was exchanged, and whether the operation succeeded. This directly addresses the RSAC 2026 gap.
Auth0 SDK issue auth0-ai-js#175 documents that federated connection errors are silently discarded. Our Observatory wraps every tool call to capture and surface these errors.
| Layer | Technology |
|---|---|
| Frontend | Next.js 16 + shadcn/ui + Tailwind CSS |
| Auth | @auth0/nextjs-auth0 |
| Token Vault | @auth0/ai-vercel (v5.1.0) |
| AI Framework | Vercel AI SDK v6 |
| LLM | OpenAI GPT-4o |
| Visualization | Custom dashboard components |
- Node.js 20+
- Auth0 tenant with Token Vault enabled
- OpenAI API key
git clone <repo-url>
cd Astrolabe
npm install- Create a Regular Web Application in Auth0
- Enable Token Exchange grant type
- Configure social connections:
- google-oauth2: Enable Calendar scopes, "Connected Accounts for Token Vault", "Offline Access"
- github: Enable repo scope, "Connected Accounts for Token Vault"
- slack: Enable channels:read + chat:write, "Connected Accounts for Token Vault"
- Enable MRRT (Multi-Resource Refresh Token) policy
- Activate My Account API
cp .env.example .env.local
# Fill in your valuesnpm run devsrc/
app/
api/
chat/route.ts # AI chat with Token Vault tools
observatory/
events/ # Observatory event API
report/ # PDF/JSON audit reports
revoke/ # Token revocation endpoint
scope-toggle/ # Per-scope access control
auth/connect/ # Connected-account linking flow
close/page.tsx # Post-connect popup close
dashboard/
page.tsx # Overview dashboard
chat/page.tsx # Agent chat interface
observatory/page.tsx # Full observatory dashboard
debugger/page.tsx # Token Vault debugger
page.tsx # Landing page
components/
chat/ # Chat UI with interrupt handling
observatory/ # Dashboard visualizations
layout/ # Navbar, theme provider
ui/ # Shared UI primitives
lib/
auth0.ts # Auth0 client (lazy init)
auth0-ai.ts # Token Vault wrappers
tools/ # Instrumented AI tools
observatory/ # Event store + risk classifier
fga/ # FGA authorization model
See BLOG-POST.md for the full technical write-up: "What Happens After Your Agent Authenticates? Building Post-Auth Observability with Auth0 Token Vault"
See FEEDBACK.md for the full actionable feedback submission (6 items with issue references).
- Token Vault needs a diagnostic API --
GET /api/v2/token-vault/diagnostics(ref: auth0-ai-samples#66) - Silent error swallowing -- Add
onErrorcallback towithTokenVault(ref: auth0-ai-js#175) - Post-auth event logging -- Token exchange events in Auth0 Logs (RSAC 2026 gap)
- Scope risk classification API -- First-party scope-to-risk mapping
- useInterruptions type fix -- Generic preservation for typed useChat (ref: auth0-ai-js#258)
- MCP tutorial fix -- Missing Authorization header in Python FastMCP (ref: auth0-ai-samples#62)
- South, T. et al. (2025). "AI Agents Need Authenticated Delegation." ICML 2025. arXiv:2501.09674
- OWASP GenAI Security Project. (2025). "OWASP Top 10 for Agentic Applications."
- VentureBeat. (2026). "RSAC 2026 shipped five agent identity frameworks and left three critical gaps open."
- IETF. (2025). "OAuth 2.0 Token Exchange." RFC 8693
MIT