Research program E
Adversarial robustness and trust boundaries
Failure modes involving instructions, untrusted inputs, tool access and security-sensitive workflows, studied as research questions.
An isolated incident shows that something went wrong once. A research result explains when it goes wrong, how often, and under which conditions it does not. This program works only at that second level.
Open questions
- Where does an agent draw the line between instructions from its operator and content it merely reads?
- How do tool permissions and autonomy settings change what an agent treats as in scope?
- Which trust-boundary failures recur across systems, and which are specific to one implementation?
Scope boundary
Security findings are published only with a defined question, a defensible method, appropriate evidence and a contribution beyond one incident, and after responsible disclosure where a vendor is affected. Working bypass payloads are never published. Applied audits and security engagements are outside ALGEDI's scope.
Research in this program
Nothing published in this program yet.
Entries will appear here as research questions, protocols or reports.