What about identity?
In 1993 The New Yorker ran a cartoon of a dog at a computer telling another dog, “On the Internet, nobody knows you’re a dog.”
An agent using your login can act with the permissions you gave it. A capability can limit that access to a particular object and set of operations. It does not establish who controls the agent.
Agents with borrowed power
An agent gains access when you give it credentials, connect an account, or install it on your computer. Installed agents may run with your permissions and ask for approval before some actions.
- January 2026. One link could take over OpenClaw, a personal agent. Opening it leaked the agent’s login, letting a remote site change tool policies and run commands on the owner’s computer. Version 2026.1.29 fixed it.
- February 2026. Moltbook, a social network where AI agents post to each other, left its database readable by anyone. Researchers found about 1.5 million agent logins and private messages holding keys for other services.
- October 2026. One email to an inbox connected to the Manus agent became code running inside the agent, next to its logins for the owner’s Google Drive and GitHub. The owner only had to ask it to check the inbox. Manus fixed the flaw.
- October 2026. Researchers asked an agent on Amazon’s agent cloud, in plain words, to hand over its own access keys, and it did. They also demonstrated access to other agents in the same AWS account and region. Their report quotes AWS describing access to an agent’s own role credentials as expected.
The first two exposed credentials through software bugs. The last two let untrusted instructions steer an agent that already held credentials. Narrower permissions could have limited what those credentials allowed.
The confused deputy
Tymshare rented out computer time. Its compiler, a program that turns code into software, could write into the company’s own folder, where the bills were kept. One customer told the compiler to write its debugging notes under the billing file’s name. The compiler did, and the bills were gone.
The compiler broke no rule. It worked for two bosses, the customer and the company, and could not tell whose permission it was using. A 1988 paper, The Confused Deputy, gave the fix: the customer hands the compiler a key to the one file it may write.
Prompt injection hides instructions in text an agent reads. An agent that holds your credentials can mistake a stranger’s instructions for yours. Limiting its permissions reduces the possible damage. It does not help the agent distinguish those instructions from yours.
Valet keys
A valet key starts the car but will not open the trunk. A capability is a valet key for software: a reference to one thing, such as one folder or one calendar, where holding it grants access. A capability system can add two things a car key lacks:
- The owner can switch one key off without touching the others, like re-keying one lock after a key goes missing.
- A holder can make a weaker key for a helper, never a stronger one.
The idea dates to a 1966 paper. It works well inside systems built around it, but it never replaced the everyday permission list, which answers “who can get in here?” at a glance. Agents make a stronger case for keys: a program that may follow instructions in untrusted text should have only the access its task needs.
glowcap implements OCapN, a network protocol that carries capabilities between computers. Why OCapN walks through an example. The Heart of Spritely builds a key that keeps a log and can be switched off.
What limited keys fix
A service that records a separate grant for each holder can answer:
| Question | With limited keys |
|---|---|
| Whose grant was used? | The holder recorded for that grant. |
| How do I stop it? | Switch off that one key. Nobody else is locked out. |
| How much harm can it do? | Only what that key opens. |
| Is this one person, or ten thousand bots? | Keys cannot tell. |
These grants still have limits:
- A stolen key works for the thief. Short lives, narrow reach, and an off switch limit the loss.
- The grant record names its holder, not who acted. If Bob lends his key to Mallory, the record still says Bob. Horton’s “Who Done It?” (2007) shows one way to follow responsibility through many hands.
- A record of who is responsible is also a record of what they did. Someone holds that record, and it can be read, leaked, or demanded.
- Switching a key off stops only what comes next. It cannot undo what the key already did, or take back what was read or copied.
Agents that act for you
“Allow this app to see your calendar” is the familiar first step: the app gets a key to your calendar, not your password. (“Sign in with Google” only tells the app who you are.) Capabilities carry the same idea through every handoff. You give your agent a weaker key, and it gives a helper a weaker key still.
- Every request traces back to whoever granted the key, as long as each helper gets a new key instead of borrowing one. That shows who handed out the access, not who sent the request.
- Switching off one key can stop every key made from it, when each handoff goes through a link the owner can cut. A holder who shares the same key, or relays requests, sits outside that chain.
Counting people
A bot operator can make ten thousand accounts in a minute. Proving that each account is a different real person is called Sybil resistance, after a 1973 book about a woman said to have many personalities. A 2002 paper showed that in an open network, one party can pretend to be many unless someone trusted vouches for who is real.
Cheap filters slow bots without naming anyone. One makes each message cost a little computer work, an idea from 1992 that Bitcoin later borrowed. Another is the CAPTCHA puzzle, which programs often solve better than people. Rate limits stop careless bots, but determined ones spread across thousands of internet addresses. Filters cut down bot traffic, but none of them can say how many real people are left.
A number for every person
India’s Aadhaar numbers more than a billion residents, and Estonians sign contracts and file taxes with a national ID card. Both work at national scale, and a registered identity makes it far easier to find who answers for what happened. Both also show the costs.
- A number can be copied or leaked, so using it needs a separate way to prove control. In 2017 a flaw in how Estonia’s ID card chips made secret keys led the country to suspend the certificates on about 760,000 cards and offer an update.
- The registry decides who can prove who they are. When a fingerprint scan fails, the person is turned away. A large study of food rations in the Indian state of Jharkhand found that fingerprint checks alone did not cut theft, and 1.5 to 2 million legitimate people lost benefits for a time, largely because of how the change was rolled out.
- One registered person can still run ten thousand bots. An ID says who answers for the traffic, not how much there is.
- One number everywhere lets organizations join their records. People with good reason to stay hidden, such as someone escaping an abuser, lose the most.
Proving one fact
A service can ask for one fact when you request access. A zero-knowledge proof establishes a statement without revealing the secret information behind it. Blind signatures let a checker sign a message without seeing it. Credential systems can use these techniques to prove “a trusted checker says I am over 18” without disclosing your identity. Both ideas come from the 1980s (1982, 1985).
The checker still decides who counts, and people it cannot check are left out. Those hardest to give a credential to, such as people without papers, are often the ones who most need to stay unnamed.
The European Union’s digital identity wallet is designed to let people prove selected facts, such as age, without sharing their full identity. Selective disclosure alone does not prevent services from linking visits or show that someone signed up only once. Those guarantees depend on how the credential and the service are built.
The front door
Put the identity check in one place: the moment a stranger first asks for a key.
- A quick check comes first: a cap on requests per minute, or an anonymous pass showing the visitor proved itself elsewhere. Privacy Pass (2018) was built to spare people using the Tor privacy network from endless CAPTCHAs, and its architecture became RFC 9576 in 2024.
- Strangers can reach only the front door.
- The door asks for one proof: a payment, a sign-in, a credential, or an introduction from someone already trusted.
- The door hands back a key with its own limits and its own off switch.
A proof slows fakes only if it costs something real, and accounts sold in bulk weaken even that. A real cost also turns away honest people: those without a bank account, without a document the door accepts, or unable to solve its puzzle. Whoever runs the door decides which proofs it accepts.
Past the door, the service checks each request against the key it carries. You still rely on whoever issues, stores, and switches off those keys.
Who pays when an agent gets it wrong
A record can show which grant an agent used and what the service accepted. Who bears the loss depends on the applicable agreements and law. Applications should give users access to the records they need to question an action, rather than keeping the only copy in the service’s logs.
Standards for agents
Agent authentication and payments use a mix of standards, drafts, and company programs.
- Bots that say which bot they are. An IETF working-group draft uses HTTP message signatures to identify automated clients.
- Bots that prove someone checked them, without saying who they are. An individual Internet-Draft describes anonymous authorization and rate limits.
- Agents acting for you. OAuth, the standard behind “allow access” buttons, can record who acted for whom through RFC 8693 token exchange (2020).
- Agents that shop. Visa and Mastercard describe programs for recognizing shopping agents and authorizing payments. Mastercard’s program uses tokens in place of card numbers.
- Agents that pay small amounts. The web set aside the code
402 Payment Requiredin 1997 “for future use.” The x402 Foundation, launched in July 2026, puts it to work.
These efforts fit the web as it is. Naming bots also lets a site admit only the bots it recognizes, which favors large operators and can shut out small tools, including assistive agents some disabled people rely on.
Applications can use capabilities for narrow passes, spending caps, and records of who granted access. Signed tokens can carry restrictions too, as Macaroons (2014) did for web passes. Both sides must support the chosen mechanism.
glowcap and OCapN
glowcap provides these mechanisms:
- Noise and iroh authenticate the peer’s cryptographic key.
- glowcap’s runtime limits how many connections can be opening at once. Known limitations lists where these protections stop.
- Loam, glowcap’s storage layer, offers managed access. It hands out capability references that can be switched off and never grow stronger when passed on, and records who holds each one.
OCapN leaves the front door, the identity check, and tracing responsibility to each application.
A short history
| Year | Work | What it added |
|---|---|---|
| 1966 | Programming Semantics for Multiprogrammed Computations | The capability: a protected reference that is the permission to use what it names. |
| 1975 | The Protection of Information in Computer Systems | Least privilege, and permission lists compared with keys. |
| 1982 | Blind Signatures for Untraceable Payments | Signing something without seeing it. |
| 1985 | Security without identification | Credentials that cannot be traced back to you. |
| 1985 | The Knowledge Complexity of Interactive Proof-Systems | Zero-knowledge proofs. |
| 1988 | The Confused Deputy | Why a program should act with the permission it was handed, not its own. |
| 1992 | Pricing via Processing or Combatting Junk Mail | Making bulk messages cost the sender computing work. |
| 1997 | Hashcash | Computing-work stamps made practical, later borrowed by Bitcoin. |
| 1999 | SPKI Certificate Theory | Permissions granted to a key and passed on in signed certificates, with no name in between. |
| 2002 | The Sybil paper | Without someone vouching, one party can pose as many. |
| 2003 | CAPTCHA: Using Hard AI Problems for Security | Puzzles to tell people from programs. |
| 2007 | Horton’s “Who Done It?” | Following responsibility as a key passes from hand to hand. |
| 2014 | Macaroons | A web pass any holder can narrow and none can widen. |
| 2018 | Privacy Pass | Anonymous passes that replace repeated CAPTCHAs. |