Exciton Research

Research

Our lead question is how to establish what an AI agent accomplished, and whether the evaluation remains reliable as agents and tasks change. Explore the questions, current writing, and next steps for each direction below.

Evaluation of AI agents

An agent can report that the work is done when the record says otherwise. The question is which observations establish what the agent changed, and whether that measurement still holds when the agent, the judge, or the task changes.

Which observations establish each requirement, and which important facts stay unobserved.
Reliability across repeated trials, with cost beside accuracy, not a single lucky success.
Whether an automated judge still grades fairly after the agent is upgraded.
A proposed study: under a fixed budget, can an evaluator choose which extra evidence to inspect and compare two agents more accurately?

Working survey, October 2026. The comparison study is proposed. No new experiments have been run.

Next step: Define the evidence-selection policy and matched-budget comparisons before testing whether they improve decisions about agent upgrades.

Read the essay

Security and privacy of personal AI agents

When an agent reads, remembers, and acts for a person, a capability can cross a trust boundary. This direction surveys how those failures happen, what an attacker must already have, and which defenses are actually evidenced.

Capability, trust boundary, failure, and impact, kept separate from attacker prerequisites and defenses.
Laboratory demonstrations, disclosed production vulnerabilities, product descriptions, and hypothetical scenarios are different kinds of evidence.
The survey does not rank products, and it does not treat a capability as a vulnerability.

Working survey. No new experiments have been run.

Next step: Extend the evidence review and identify which failure mechanisms have been independently reproduced.

Read the essay