Research
A few problems I keep coming back to, what I've shipped on each, and what I still don't know how to answer.
Accountability for AI agents
Agents now write code, call APIs and act for people. When someone asks what an agent did, the answer usually comes from logs held by whoever ran it. With Euthyne I'm working on records an outsider can check: what the agent was allowed to do when it started, and a signed receipt for each action, marked inside, outside or unknown against those permissions. Content is stored as salted hashes, receipts go into an append-only public log anchored with OpenTimestamps, and anyone can verify them offline. Coding agents come first, because that's where the question already gets asked, in security reviews and audits. The signing is the easy part. The hard part is deciding what a record can honestly prove and being clear about what it can't.
Trust registries and wallet interoperability
Identity wallets decide who to trust from a list baked in at build time. At 2060.io I work on Verana, a public registry of who may issue or verify which credentials. I wrote the OpenID4VC issuer and verifier in VS Agent, which refuses to proceed when the registry doesn't vouch for the other side, and I patched existing wallets to do the same lookup. I co-authored the Verana trust specifications. The preprint Conformant but Not Compatible came out of testing thirteen production wallets against each other: they pass their own conformance tests and still can't talk to one another.
How AI engines read the web
AI engines now describe companies to people who never visit their site, and nobody at the company is asked. At CiteMe we measure that answer and the traffic behind it. Two studies so far. The first looked at 385,057 AI-agent requests across 20 sites and found that 7% of what passes for AI crawler traffic is something else. The second shows you can tell which AI assistant fetched a page from its network calls alone, but you can only tie that to a visit when the assistant reads the page live.
Open Questions
- 1.When an agent's permissions change mid-session, which version was it really acting under, and how do you prove it after the fact?
- 2.How much of an agent's work can you prove to an outsider without showing them the code it touched?
- 3.A live trust lookup is a network dependency. What should a wallet do when the registry is down, and does fail-closed survive real users?
- 4.Can a model ever check who is behind a claim before it repeats it about a company?