Research topic
Agent Security and Control
Research on trust boundaries, agent permissions, evaluation, privacy, cybersecurity, and the controls required when software can take consequential actions.
Working answer
How can autonomous systems remain useful, observable, and bounded?
Useful autonomy requires explicit boundaries: least-privilege tools, inspectable state, independent evaluation, reversible actions, and escalation paths that survive real operating pressure.
Signals this trail tracks
- Agent risk is increasingly an execution-system problem, not only a model-safety problem.
- Memory and tool permissions create durable trust boundaries that need separate review.
- Evaluation must test consequential workflows and recovery behavior, not just answer quality.
Source-backed briefs