DSS 2026 / Day 2 / Amber Hall
Your laptop is in scope too
Jan Kalivoda
Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.
Transcript
0:00My talk is about personal security. So
0:04when you are developing something or
0:06auditing, so you don't get hacked during
0:08that. This was initially supposed to be
0:12a workshop, but
0:14changed the talk, so it will be covered
0:16just briefly. But definitely you can
0:19then at your hotel apply all these
0:21things, so you will be sure that you
0:23will keep secure.
0:25So why is this important?
0:28You probably noticed also
0:31few days ago the Binance hack.
0:33So it was kind of severe and so from
0:37what I heard it also somehow involved
0:39compromised developer machine.
0:42And these attacks are more and more
0:44common.
0:46You are you are now targeted anywhere,
0:48so you should be
0:51more focused.
0:54Yeah.
0:56Today we will cover four topics.
0:59We will start with IDE, package
1:01managers, GitHub, and your AI coding
1:03tools. So the most important stack to
1:07create or audit something I will say.
1:10So starting with IDE,
1:12you probably know that you have there in
1:14VS Code or Cursor that JSON setting.
1:18There are just
1:19few of them.
1:20First one is the workspace trust, which
1:23is
1:24recommended to be turned on because when
1:27you open
1:28some
1:30some folder,
1:31it will prevent execution within that.
1:34So you just need to also click a button
1:38to
1:39trust it.
1:40You can also turn off automatic tasks,
1:45which is kind of recommended. Is there
1:47anyone who is using
1:49this task in VS Code?
1:53Did not encounter anyone yet.
1:56So
1:57you should have this definitely
2:00We will show later why. And then I
2:04personally like
2:06when when I when I have control over
2:09updating some packages and essentially
2:12all the software. So I'm just turning
2:14off all the updates
2:16and checking that always in my weekend
2:19sessions manually
2:21that I'm not pulling
2:23any malicious version.
2:27This is how the
2:29workspace
2:30window looks like.
2:32And I'm not sure if it is clearly
2:35visible, but this is example project
2:37which can trigger arbitrary script on
2:40your computer.
2:42You basically open just some folder and
2:45it has .vscode folder there and there is
2:49defined a task.
2:50So and this task
2:53is executing harmless.
2:57SH
2:58file
2:59and as you can see there is a lot of
3:03other arguments that are passed
3:05and those arguments are essentially
3:07doing that it will be executed very
3:09silently, so you won't notice that.
3:12So
3:14when you were there nothing happened and
3:17when you click trust
3:19then after 30 seconds, which is
3:22the script, you got just executed it and
3:25there there was created a file.
3:32Then definitely
3:34problem is with extensions. So
3:39what I recommend for companies, you can
3:41create
3:42approved list of extensions where you
3:44pin to exactly some version and you can
3:47distribute it between all your employees
3:50so you will ensure that no one will be
3:52using any unapproved extension
3:56in their editor.
3:58And
3:59reviewing your updates and also you may
4:01be
4:03someday in the past, you know, you have
4:05VS Code probably installed for years,
4:07and it's still evolving, it's having new
4:09parameters and so on. You may be click
4:11there that you have a full disk access
4:14or other things.
4:15So it's also good to review this.
4:17Maybe you don't need it for that.
4:21Then we have their package managers.
4:24So
4:26the flow basically for those basic
4:28supply chain attacks is that there is
4:31some maintainer of some package or his
4:33CI gets compromised,
4:35update is released,
4:37and
4:40you will just execute it on your machine
4:43and
4:44get for example credentials stolen.
4:52So some basic settings what you should
4:54keep in mind,
4:56have scripts off. It's basically similar
4:58to the thing we were describing in VS
5:01Code.
5:03And
5:05also in VS Code, and I I'm not sure if
5:07in VS Code itself, but I know that there
5:10was that option
5:12in VS Code.
5:13You can set there a freshness gate for
5:17some packages, and this should be also
5:20definitely set
5:22in your package manager setting.
5:25Then
5:26as always you have to check for
5:28overrides and that your toolchain is
5:30properly pinned.
5:33And this is kind of funny because you
5:36have a lot of package managers and they
5:40are setting those things
5:42pretty differently.
5:43So let's say you will just check it,
5:47set it correctly for npm,
5:49let's say also for pnpm,
5:51but then you will just change version
5:54with nvm, let's say, and then you will
5:57be using completely different config
5:59setting,
6:00which is not the one that you you were
6:02supposed to use. So
6:05again scripts can get executed. So, you
6:08have to have it like covered uh
6:12in all many possible ways
6:15or you have to ensure that you will be
6:16always using just the version you want
6:18to and where you have the config
6:20specified.
6:22Uh this is example for uh
6:25the gates.
6:27Uh I personally like most PNPM. Uh I
6:31think they have the best attitude to
6:34security. So, I'm also enforcing it to
6:37use it by my agents and always their
6:39specific version range.
6:43So, then we have a GitHub.
6:45Uh
6:46who is using GitHub CLI?
6:49Here.
6:50One? Just one person? I was expecting
6:52more.
6:54Well, um
6:55I stopped using that a year ago because
6:59I realized that by by default that tool
7:03was saving
7:05uh
7:06your access key in plain text. And when
7:08you removed the GitHub CLI tool, then uh
7:12it's still there was the plain text
7:15existing.
7:16So, if you are using GitHub CLI, then I
7:18urge you
7:20to check this because you should not
7:22have any secrets in plain text on your
7:24machine.
7:26And
7:28then specifically for the GitHub,
7:32uh
7:33least privilege
7:35approach is the viable way everywhere.
7:40And also for Git, you have to be aware
7:43of Git hooks. So, you should disable
7:46them.
7:47And more importantly, if you are for
7:49example auditor and you will get a code
7:52base from your client and you want to
7:54push it to your let's say internal
7:56organization to collaborate there with
7:58your
7:59colleagues, then you have to have keep
8:02in mind to the actions because
8:05there can be anything even though it is
8:07your client, Uh
8:09but you can potentially trigger some
8:11malicious action.
8:13So
8:14what setting would I like
8:17is Is it a laser?
8:20No.
8:22Is the last one
8:24where you essentially approve all your
8:26actions what you created. You are
8:28creator, so you probably don't need to
8:31be so much aware
8:32if it's simple
8:34or like that.
8:36But
8:37you can also select some specific
8:40actions which you don't maintain and you
8:43can pin them to specific hash
8:45and this way you can just keep again
8:48list similar like with extensions in VS
8:51Code
8:53of actions what what you
8:56what to use.
8:59And then definitely repository controls
9:05are important to be set properly.
9:08Then we have there section with AI
9:11coding agents.
9:13There I will have again a question.
9:16Is there someone using auto approval
9:18mode?
9:22And someone clicking allow?
9:26Okay.
9:27I think
9:29if you will be at this state like auto
9:31approval if it will be configured good
9:35it can be actually somehow useful, but I
9:38think it will be still stopping you in
9:39some bigger progress and I personally
9:42don't trust it.
9:44So when you're running agents I think
9:47it's just a matter of time when you will
9:49start with proper sandboxing.
9:52So
9:55I will recommend to start with
9:57containers.
10:00For example,
10:02I'm not using Cloud Code or Codex. I
10:03like Ollama Pi harness and it's by
10:06default like YOLO. So
10:09you can't even run it safely on your
10:11machine. So this was also like my
10:13motivation
10:15to have it set properly.
10:17But
10:19even to even you are running it in
10:21container or outside,
10:23you can still
10:24nicely configure these tool tools.
10:27So let's say you can
10:30create a recoverable cleanup.
10:32You can replace
10:35remove calls with trash command.
10:38This can be little bit tricky because if
10:41you are not cleaning your trash
10:42regularly, then it will kind of grow.
10:47Then you can have a specific
10:50rules for feature branches and work with
10:53Git.
10:54This can be also stacked in the layers,
10:57one layer on directly on GitHub and
10:59another on the agent level. And
11:02so you will ensure that
11:04you have stacked up the security in
11:05layers.
11:06And
11:08yeah, and so many other things. There
11:11are examples how these configs can look
11:12like.
11:15You can deny some folders, do lots of
11:18stuff.
11:20For codex, it's in this way.
11:23And
11:26what I
11:28would like you to do when you will be
11:31coding with AI agents to have set the
11:34sandbox. So for example, what works
11:37nicely for me are just simple dev
11:39containers when you can have for example
11:41dev container for creating dev
11:42containers and then you can create dev
11:44containers for each task.
11:46Mount always just the folder what you
11:48need.
11:49And
11:51then you can run your agents and box in
11:54this way while still they have access to
11:56your laptop or if you are running it
11:59remotely, that's even better when you
12:00have a dedicated machine
12:03for the coding.
12:07Yep. And that's the talk.
12:11Questions.
12:13>> So when you're building um
12:16that Docker image, how do you make sure
12:19that the agent can escape and come back
12:22to your, you know, laptop and actually
12:25pull secrets off?
12:27>> That's a very good question. Um
12:30I'm not sure. Uh that's why still the
12:33dedicated machine will be better.
12:35But uh
12:36uh I think
12:38the versions of agents we are using uh
12:41are much more constrained than than
12:42those in some
12:44lab scenarios where uh
12:47where the frontier labs are are actually
12:50testing them. So, they have them little
12:52bit like unchained. Meanwhile,
12:55we have still there lot of
12:58guard rails and so on. So, I would
13:00suppose that uh
13:02like I think in security there is always
13:04way
13:05how you can hack yourself or someone
13:07else.
13:08But
13:09uh it's generally about trying to do
13:12most of the mitigation you can do and
13:14still
13:15preserve some comfortable way of being.
Source and accuracy details
- Names, organizations, technical terms, code and numbers may be misrecognized.
- Source wording and disfluencies are preserved; no editorial corrections were applied.
- Export contains whole-second cue starts only; cue ends and subsecond timing are unavailable.
- A full caption-track export does not establish that every spoken word is represented.
- Automatic YouTube captions; not human-verified against the audio.
Source export SHA256: c58dc59396f77611567ce0d81489071edd4a29198d742610814e87cf83302b61