Agent identity and delegated access: how AI agents act for people
10
Glossary and sources
5 min read · Chapter 10 of 10
Words that mean something else here
| Term | What you'd assume | What it means here |
|---|---|---|
| Agent | One kind of thing | An OAuth client, a new identity type in Entra, a user-triggered fetcher to a website, a registered actor to a card network |
| Delegation | Handing a task over | In token exchange, the agent keeps its own identity beside the person's; impersonation is acting as the person |
| Consent | The user clicked Allow | An OAuth screen, an admin's approval at the identity provider, a CIBA approval on a phone, or the agent's own permission prompt |
| Verified (Cloudflare) | Allowed in | The agent is who it says it is; since July 2026, nothing more |
| Authorized | Allowed by the platform | In payments rules, done by someone with the consumer's authority; in Amazon's case against Perplexity, permitted by the site |
| Confused deputy | A vague security worry | In the MCP spec, a precise attack on proxies with a shared client ID; elsewhere, any injection that turns an agent's own access against its user |
| Draft | Nearly done | An IETF Internet-Draft, which expires in six months unless revised |
| Immaterial | Zero | Okta's word for agent revenue this fiscal year: too small to report |
What surprised me
Placeholders in your voice, drafted from the research and the earlier guides. Rewrite each with your own moment.
"Fake agents are the threat." I thought spoofing would be the story, since a user-agent string is so easy to copy. Every incident with a real loss used a real agent or a real token.
"Verified means let in." Coming from the login guide, a verified identity felt like the end of the check. Cloudflare split the two in July 2026, and the site's rules now decide.
"Asking the user keeps it safe." People approve 93% of prompts in the one dataset I found, so the prompt mostly records consent after the fact.
"Someone has decided who pays." In payments, every failure had an owner in the rulebook. Here nobody has written one yet.
Sources
Undated entries were read on October 13, 2026; "search result" means seen only as a search snippet or summary. Company figures are self-reported unless they come from a filing or a regulator, vendor surveys cover the vendor's own respondents, and list prices change often.
Standards and specifications
- Model Context Protocol: authorization specification (Jul 2026); changelog; security best practices; enterprise-managed authorization (Jun 2026); SEP-1932, DPoP profile
- IETF: OAuth 2.1 (draft 16, Sep 2026); ID-JAG (draft 04, May 2026); Client ID Metadata Document; AI agents on behalf of users (expired); Web Bot Auth working group and protocol draft (Sep 2026)
- RFC Editor: RFC 9728 (Apr 2025); RFC 9396 (May 2023); OpenID Foundation, CIBA Core 1.0 and whitepaper on agentic AI identity (Oct 2025)
- Visa, Trusted Agent Protocol specifications; AP2; FIDO Alliance, agentic commerce standards (Apr 2026, search result); Solo.io, MCP authorization critique (Aug 2025, vendor)
Rules, guidance and courts
- NIST: IR 8587 (Sep 2026); NCCoE concept paper (Feb 2026) and project page
- CISA, joint agentic AI guidance bulletin (May 2026); OWASP agentic Top 10 via Giskard (search result)
- EU: AI Act Article 50; Gibson Dunn, AI Act omnibus (search result); Taylor Wessing, product liability for AI (search result); Fintech Global, eIDAS 2.0 deadlines (search result)
- Payments: CFPB, Reg E 1005.2 and Reg Z 1026.12; Hogan Lovells, PSD3 and PSR (search result); Linklaters, UK payments priorities; Bratby Law, agentic payments and consent
- Future of Privacy Forum, who pays when transactional agents play (Feb 2026); Baker McKenzie, legal accountability for AI agents
- Amazon v. Perplexity: PPC Land, injunction (Mar 2026); Retail Insight Network, injunction vacated (Aug 2026)
Terms and insurance
- Anthropic, consumer terms and commercial terms; Google, Gemini API additional terms; Microsoft, Services Agreement; Perplexity, terms of service and Comet terms; Replit, terms of service
- Worldpay, agentic commerce liability (vendor); American Express, agentic commerce announcement (Apr 2026, search result)
- Insurance Journal, AI exclusions (Aug 2026); Armilla, $25 million cover (search result, vendor); SiliconANGLE, AIUC raises $40 million (search result)
Incidents and security research
- Drift: Google Threat Intelligence, Salesforce data theft (Aug 2025); Cloudflare, its Drift response (Sep 2025); The Register, Salesforce class actions (Sep 2025)
- Invariant Labs, GitHub MCP (May 2025); Simon Willison, Supabase MCP (Jul 2025); Supabase, defense in depth for MCP (Sep 2025); The Register, Asana MCP bug (Jun 2025)
- MCP Inspector advisory (Jun 2025); The Hacker News, Git MCP server flaws (search result); EchoLeak case study (Sep 2025); Zenity, AgentFlayer (Aug 2025); Petri, Copilot audit-log fix (search result); TechRadar, Antigravity (search result)
- The Register, Replit and Replit's response (Jul 2025); AWS, AWS-2025-015 (Jul 2025); The Register, postmark-mcp (Sep 2025)
- Trend Micro, Vercel breach (Apr 2026); CSO Online, Klue breach (Jun 2026)
- Anthropic, Claude Code auto mode (Mar 2026); GitGuardian, State of Secrets Sprawl 2026 (vendor)
Agents on the open web
- Cloudflare docs: Web Bot Auth (updated Oct 2026), verified bots, signed agents, pay per crawl (updated Jul 2026)
- Cloudflare blog: Web Bot Auth (May 2025), signed agents (Aug 2025), pay per crawl (Jul 2025), Perplexity's undeclared crawlers (Aug 2025); Help Net Security, Cloudflare's crawler controls (Jul 2026)
- Crawler pages: OpenAI; Anthropic; Google, Web Bot Auth (updated May 2026) and user-triggered fetchers; Perplexity; Simon Willison, ChatGPT agent's user agent (Aug 2025)
- Akamai, bot management in the agentic era (Nov 2025); HUMAN, Atlas and Comet (search result, vendor); DataDome, Web Bot Auth and Know Your Agent (search result, vendor)
- Visa, Trusted Agent Protocol (Oct 2025) and Visa Europe agent transactions (Jul 2026, search result); Mastercard, Agent Pay (Apr 2025, search result)
Companies, products and pricing
- Okta: Cross App Access (Jun 2025); Agent SSO (Aug 2026); Global CISO Insights 2026 (vendor survey, n=306); Futurum, Okta's Q2 call (Aug 2026)
- Auth0, AI agents generally available (Nov 2025) and pricing; Microsoft, Entra Agent ID, what's new, key concepts, Entra pricing and message MC1395007 (mirror); Rencore, Agent 365 pricing (reseller, search result)
- AWS: AgentCore Identity, AgentCore pricing, AgentCore Browser signing, Cognito pricing; Google Cloud: Agent Identity, IAM release notes, Next '26; Anthropic, Managed Agents vaults (search result)
- Pricing pages: Clerk, Descope, Arcade, Cloudflare plans; Vendr, DataDome purchases (search result); Kasada, Forrester Wave for bot and agent trust (search result); Silicon UK, Gartner on cancelled agent projects (search result)
Deals and funding
- Security Boulevard, identity market after consolidation (Oct 2026, vendor-authored); Cisco, Astrix (search result); CrowdStrike, SGNL (Jan 2026, search result)
Reused from earlier guides on this site
- OAuth and token basics, Stytch's sale to Twilio, Okta's and Clerk's liability caps, and the Entro, Permiso and CyberArk deals: Login and access for apps; company-size budget bands: Cloud security (CNAPP); prompt injection and MCP tool attacks: AI agents: orchestration platforms
Field Guides are learning notes, not legal or compliance advice. Rules and fees change; check the cited primary sources before you act on anything here.