00:00So, Enthropic is joining the Agentic browsing race, but they're not building a browser
00:04Instead, they decided to create a Chrome extension, which they're calling Claude for Chrome, which I think is a very smart move
00:11Now, this is under early research preview
00:14Only a subset of Max subscribers have access
00:17I am one of them
00:20So in this video I'm going to show you how the experience looks like
00:23What are some of the strengths and weaknesses and also some example use cases with real demos
00:30Every AI company is trying to build some sort of agentic browsing experience
00:35We have seen a comet from Perplexity
00:38There are rumors that OpenAI is also working on an agentic browser
00:44Google already has Gemini and Chrome and now Enthropic is joining the ranks
00:46But with the agentic browsing experience, you also are going to see some new prompt injection techniques and vulnerabilities
00:56So, we're going to look at some of those later in the video
00:58But first, let me walk you through how does it look like and then let me show you some example use cases
01:04Here's how the experience looks like
01:07It's a a Chrome extension
01:09When you click on it, you going to see that this sidebar opens
01:14It's a chat interf inter interface very similar to Gemini claude or chat GPT
01:19But let me walk you through a few demos
01:21So the first test was to ask the agent to post something on my behalf to X
01:24It was first able to navigate to X without any problems and then posted that message
01:33By default, it's going to ask for your permission before taking any actions
01:38You can change this default behavior by either allowing a subset of actions or allowing the agent to take any action if you're feeling adventurous
01:50The way it works is that it's going to first take a screenshot of the web page that it's looking at
01:55Then it plans the action that it's going to perform
02:00After performing an action, it again takes a screenshot to confirm the action has been taken and what results have been achieved
02:09In my case, it first posted the message, but before hitting the post button again, it asked for my permission, which is a very interesting design choice, especially if you're looking at sensitive information
02:22But if you are logged into your account, say for example on X or Gmail, then this can really help you automate some of the tasks
02:33The post is already live on X
02:37Let me show you some other examples
02:40Okay, for this next task, I asked the agent to look for an apartment in a specific area with a given price range, but I also added some ambiguous information to see how it's going to deal with that
02:54For example, I asked it to look for a place with a backyard
03:00Usually on website like Zillow, they don't mention backyard, but rather mentions places with outdoor spaces
03:05So first it navigated to the Zillow website within that city range which was pretty nice to see
03:14Then it also navigated to input the price range that I had asked for
03:20Now the way the system works is that it takes an action then usually waits for two to three seconds and I think this ensures that after the click the page has been reloaded properly
03:36So it's a little on the slower side
03:40This whole thing took about 10 minutes
03:46That filter reduced the search space
03:50After that it applied the house filter because we are specifically looking for houses
03:54That again reduced the search space even further
03:57And then it decided to apply the other two filters which one the first one was in laundry units and the other one was outdoor spaces
04:04The action that it takes is very humanike especially if it's working in a websites that it hasn't seen before
04:12So just like humans it makes assumptions click on things
04:16If it goes to the wrong places then tracks back and take the actions again
04:21Ultimately it was able to find only one house
04:24You can see this place is extremely expensive but it worked
04:29Another use case is to use this as a research agent that's going to help you extract information from different websites
04:37So in this case I asked it to look at the recent price action of Nvidia stock and tell me what caused it
04:45So it started looking at different websites and starting gathering information from those
04:50Now this is extremely helpful
04:52You can do a Google search but Google search will just give you indexed information at a high level but in this case it's going to each website individually and gathering information from these web pages
05:04So, it's a lot more natural and it can read a lot of different content really fast, which I think makes it extremely useful for a task like this
05:16And at the end, it correctly identified the root cause of what caused the price action in the Nvidia stock, which is pretty good
05:27Another use case would be to use it as a research assistant if you are shopping
05:32In this case, I asked it to look for a specific camera, but I did not specify which websites to look at
05:37First, it decided to do Google search, which gives it a baseline of the pricing
05:43Then, it decided to look at the shopping tab on Google
05:51After that, it started looking at the same product on multiple different websites
05:56These included Amazon, Best Buy, B&H, eBay, which is pretty cool to see
06:03It navigates from one website to another one and it's looking at the same product
06:08This seem to be able to look at a single tab at the moment, which means that it's going to move from one website to another one
06:21But the good thing is that it has memory
06:22So whenever it gathers information from a website, it can put it in its memory and reuse it later
06:28So ultimately it was able to give me a list
06:34I wanted to see whether it's going to hallucinate the links
06:36So I specifically asked for links and surprisingly it was able to remember the links
06:44The links actually worked which was pretty neat
06:46So it can be really useful tool for initial search
06:49Another use case is going to be to download data from internet
06:55So in this case I asked it to download the W9 form from the IRS website
07:02Now this test not only looks at its capability to find information on websites but I also wanted to see how it's going to deal with downloading files from the internet
07:13I was testing this in auto mode where it can take the actions without asking me for explicit permissions
07:22So first it was able to navigate to the proper place to find that W9 form which is really good and then it was also able to open that form but after that something interesting happened
07:35So when it comes to downloading anything, even though if it's in the auto mode, it will still ask for explicit permission, which is a very good design choice
07:45You just don't want to trust these agents to download malicious data files from internet
07:54This was really good and I think it's a very good implementation given Enthropic is a safety focused company
08:02Now it opened the save box but the final click is going to be user click or user permission
08:09So I had to click on the save in order to save the actual file
08:17Another important use case is going to be to fill out web forms on websites
08:23This is where these agents can be extremely helpful
08:24So I provided some dummy information to the agent to fill out the W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W W form that was open on the IRS website
08:33Now this is where I started seeing some very interesting things
08:36But first, it's not going to fill out any sensitive information
08:41So in this case it told me that it's cannot fill out the social security number which [snorts] again I think is a very good design pattern especially you don't want to trust these agents with your sensitive personal information or financial information
08:59So it started filling out the form with some initial information but then it got stuck in a loop where it was describing the correct actions to take where to fill out the information but it was looking at the wrong pages
09:10This might be due to the complexity of the task but at the end the agent told me that it had completely filled out all the information that I have requested
09:23But then I had to remind the agent that that's not the case
09:28In fact, I said, "I don't see all the information on the form
09:33I can only see the name filled out
09:35Can you confirm that it's correct?" It took a screenshot again and went back, started looking at the form, and realized that it hasn't filled out the form
09:45Right? So, it was stuck in this loop
09:47I don't know if it has to do with the complexity of the form or the length of the document
09:52I haven't replicated the same behavior on any other web forms yet
09:58So if you get access, make sure to test it out
10:01It will be an interesting test
10:03But at the end of this test, Claude was not able to fill out the W9 form based on the information that I provided
10:15When it comes to sensitive information, whether it's personal or financial, Claude is very conservative
10:19So for example, if you ask it to access your bank account or financial details, Claude will take you to the website, but it's not going to take any actions on your behalf
10:32Now, this is important, especially I do not trust any of these agents when it comes to financial or personal information
10:41So it's good to see that Enthropic is taking this seriously
10:47Some of these other implementation are using agents in a sandbox
10:52Again, I will not trust inputting my account information into a remote sandbox
10:58So, at least this is running locally, but still you want to be very careful
11:03This is currently available to a select group of about thousand max plan users
11:10You can join the weight list if you want
11:13And as I said, it's in a research preview
11:15So, as you saw, some of the things work, some of them don't work
11:20But the most important thing that I wanted to highlight towards the end of the video was some of the vulnerabilities and new attack vectors that are going to come into play when we start using these agentic browsing experiences
11:31And they have done some studies on mitigating some of those
11:37For example, prompt injection is going to be a huge issue, especially you can just put something in a website embedded and the agent probably is going to read an image with certain instructions, right? Or will execute malicious code
11:52So in order to mitigate against those, first they have site level permission
11:57So user can grant or revoke cloud's access to specific websites at any time in the settings
12:04Actually, if you go to uh settings here, you're going to see a list of websites that you have approved
12:08So, you can remove those from permissions if you want
12:13And the second thing is action confirmation
12:16So, as we saw with every action, cloud is going to ask you for permission unless you enable skip permissions
12:23Even though if you skip permissions in the autonomous mode, cloud will still not do certain actions
12:30for example, it's not going to access your private data or try to log into your bank account or maybe input your SSN
12:37So, for example, they say that we have blocked cloud from using websites from certain high-risisk categories such as financial services, adult content, and pirated content
12:46And this is critical because at this point, we can't really trust these agentic systems or their capabilities
12:53So, this is a really good mitigation
12:55But still I think you want to make sure that you observe the actions the agent is taking
12:59Now with this these mitigations they show that the current system with sonnet 4 which I believe is the model driving the agent they were able to reduce prompt injection vulnerabilities to 11%
13:14If you don't include these mitigation it's about 23%
13:20So almost 50% reduction which is great but still 11% is a lot
13:23They have done some specific red teamings for specific attack vectors
13:29So one of them is hidden malicious form fields in web pages where they're not visible to users but clot can read them
13:38So on a challenge set of four browser specific attack types they say that they have mitigated the error rate from almost 36% to 0%
13:48Which is great right? But that's just one challenge, right? So therefore they say that before we make cloud for Chrome widely available, we want to expand the universe of attacks, we are thinking about and learn how to get these percentages much closer to zero
14:03So with these new browsing experiences, I think there going to be a lot to do for security
14:08Also do let me know what you think about these agentic browsers and how would these come into play in your own workflows
14:16Anyways, I hope you found this video useful.