WEBVTT

00:00:00.900 --> 00:00:04.940
Before you hand an agent a file,
stop and ask where it's going.

00:00:06.360 --> 00:00:12.065
Three questions: where does it go, does
it belong there, and whose rules apply?

00:00:14.800 --> 00:00:17.748
AI for Research Efficiency.

00:00:17.798 --> 00:00:21.707
Episode 9: Your data,
your keys, your campus.

00:00:24.300 --> 00:00:28.125
Last time, many agents at once:
several terminals,

00:00:28.196 --> 00:00:30.378
or one agent that splits the work.

00:00:31.163 --> 00:00:35.608
This time, the rules around all of
them, with a checklist at the end.

00:00:38.900 --> 00:00:41.911
One: your plan decides where it goes.

00:00:42.811 --> 00:00:46.025
Whatever the agent reads goes
to the company that runs the model,

00:00:46.176 --> 00:00:47.701
along with your request.

00:00:49.222 --> 00:00:52.452
On a personal plan, it may be used
to train future models,

00:00:52.563 --> 00:00:54.427
unless you switch that setting off.

00:00:55.641 --> 00:01:00.665
On a work, school or API plan, it
isn't used for training by default.

00:01:02.254 --> 00:01:06.202
So check which plan you're on,
and its setting, before you start.

00:01:10.100 --> 00:01:12.571
Two: what never goes in.

00:01:13.295 --> 00:01:18.583
Student records, health data, study
participants' data under an IRB protocol,

00:01:19.054 --> 00:01:22.900
export-controlled work,
and other people's unpublished work.

00:01:24.140 --> 00:01:25.566
None of it goes to an agent,

00:01:25.738 --> 00:01:28.832
unless your campus has approved
that tool for that data.

00:01:30.742 --> 00:01:34.813
At the University of Arkansas, for
example, data comes in four classes:

00:01:35.185 --> 00:01:39.276
restricted, highly sensitive,
sensitive, and public.

00:01:40.816 --> 00:01:41.458
Restricted

00:01:41.508 --> 00:01:45.592
and highly sensitive data may only
go into tools the university licenses

00:01:45.642 --> 00:01:47.005
or approves for it.

00:01:50.900 --> 00:01:52.656
Three: keys.

00:01:53.477 --> 00:01:57.923
A key is a password your scripts use,
and some keys can spend money.

00:01:59.047 --> 00:02:01.253
Put it in a file named .env,

00:02:01.665 --> 00:02:06.398
and list that file in .gitignore,
so it never reaches a repository.

00:02:08.718 --> 00:02:12.226
Then give the agent the key's name,
never its value.

00:02:12.617 --> 00:02:16.524
A pasted key goes to the vendor,
and stays in the session's history.

00:02:18.068 --> 00:02:21.193
Without a rule,
an agent reads the file when you ask.

00:02:21.665 --> 00:02:24.850
Here, Codex prints the
value into the conversation.

00:02:27.135 --> 00:02:29.677
So write the rule into
your instruction file:

00:02:30.127 --> 00:02:32.359
never read or print .env.

00:02:33.050 --> 00:02:35.752
Claude Code can enforce it
with a deny rule.

00:02:39.492 --> 00:02:42.414
Codex can deny the file
in a permission profile.

00:02:42.945 --> 00:02:48.128
Antigravity's sandbox is meant to block
it; in our test, it read the file anyway.

00:02:49.848 --> 00:02:53.853
Give each project its own key,
with only the access it needs.

00:02:54.224 --> 00:02:56.667
If one leaks, revoke it first.

00:03:00.500 --> 00:03:02.936
Four: someone else's work.

00:03:03.756 --> 00:03:06.942
If you review grant proposals,
the funders have spoken.

00:03:07.453 --> 00:03:08.384
At NIH,

00:03:08.495 --> 00:03:11.931
reviewers are prohibited
from using AI tools in analyzing

00:03:11.981 --> 00:03:16.609
and critiquing NIH grant applications
and R&D contract proposals.

00:03:17.970 --> 00:03:22.217
NSF reviewers are prohibited from
uploading any content from proposals,

00:03:22.307 --> 00:03:27.455
review information and related records
to non-approved generative AI tools.

00:03:29.080 --> 00:03:31.431
A journal's peer review
has rules of its own.

00:03:31.823 --> 00:03:34.213
Read them before you
open the manuscript.

00:03:36.500 --> 00:03:39.476
Five: find your campus's rules.

00:03:40.316 --> 00:03:41.978
At the University of Arkansas,

00:03:42.269 --> 00:03:47.055
any AI tool used for university
business must be vetted and approved.

00:03:48.415 --> 00:03:50.696
Its approved list names chat tools.

00:03:51.187 --> 00:03:55.079
On 4 October, none of the three
agents in this series was on it;

00:03:55.389 --> 00:03:58.231
for those,
there's a tool exception request.

00:03:59.994 --> 00:04:01.837
Your campus will have its own list.

00:04:02.208 --> 00:04:07.298
Search its website for approved AI
tools, or ask its IT security office.

00:04:10.100 --> 00:04:12.067
Pause here, and check two things:

00:04:12.398 --> 00:04:16.271
your tool's training setting,
and your campus's approved list.

00:04:19.700 --> 00:04:23.127
So: check your plan,
keep protected data out,

00:04:23.478 --> 00:04:25.292
keep keys in .env,

00:04:25.964 --> 00:04:29.591
keep reviews out,
and find your campus's list.

00:04:33.900 --> 00:04:38.020
Next: research APIs,
a ladder of keys.