[-]

legrenabeach@reddit

I don't know the technical aspect, but I know how others handle it. A friend of mine is a developer for a major UK Bank. He works hybrid, two days from home and three days in the office. He can only use the company laptop to access anything work related, as it uses a VPN to access the bank's network and allow access to all internal tools etc, no matter what broadband or mobile data connection he uses. They use two different LLMs, one internally developed and a mainstream one with a company account. The rest is blocked, I believe both at company firewall and at local laptop level.

ElonTaco@reddit

You obviously do not understand what distilling a model is.

[-]

youplaymenot@reddit

It's not the same exact model genius, you think anthropic is just letting anyone download their exact cloud model? Let alone having massive amounts of ram to run that model.

[-]

Superb_Raccoon@reddit

https://huggingface.co/Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled

Same inference engine.

https://techcrunch.com/2025/10/07/anthropic-and-ibm-announce-strategic-partnership/

You probabl6 missed this. IBM runs Claude on prem, and provides it to clients on Prem as part of Orchestrate..

[-]

bingblangblong@reddit

Decided to look into this.

https://newsroom.ibm.com/2025-10-07-2025-ibm-and-anthropic-partner-to-advance-enterprise-software-development-with-proven-security-and-governance

There's no mention here of IBM running their own Claude, same with the TechCrunch article. It only says:

Armonk, New York-based IBM announced it will be adding Anthropic’s Claude large language model family into some of its software products on Tuesday. The first product to tap Claude will be IBM’s integrated development environment, which is already available to a select group of customers.

So what that probably means is they're using Anthropic's API to access Claude. Not the same as IBM running it themselves.

I could not find any mention anywhere of Anthropic publicly releasing a model. What that huggingface.co page says is:

The model includes targeted optimizations addressing Qwen3.5’s tendency toward excessive transitional or repetitive reasoning on simple queries. Through deep distillation and structural imitation of Claude-4.6-Opus reasoning chains

"imitation of Claude-4.6-Opus reasoning chains" so it is... not at all Opus 4.6. It is trying to be like Opus 4.6.

The "inference engine" refers to the software that runs the model, like vLLM, llama.cpp etc. That's completely separate from the model weights themselves. I think you are misusing the phrase to imply the model is the same, which it is not.

[-]

Superb_Raccoon@reddit

I work there. I know.

[-]

bingblangblong@reddit

OK I believe that you work there but you misused inference engine to imply it's the same model and the link you provided is demonstrably not Claude. You also posted the techcrunch article to validate your claim that IBM runs Claude locally but it doesn't say that so I'm not sure why you posted it.

[-]

MaNbEaRpIgSlAyA@reddit

Opus, Sonnet, and Haiku have never been open sourced.

[-]

Superb_Raccoon@reddit

So?

Who said they were?

I downloaded RAM does that count? I dont think it was made by dodge though...

[-]

UMustBeNooHere@reddit

gif

You can run smaller deepseek models with 8B parameters with 12gb gram, and bigger 32B parameter models with 24gb of vram, so an rtx 4090. Most use cases really don't need a 670B parameter model.

If you're a smaller company that for some reason rally needs a private LLM you could be done with an about 5k USD investment. And if you really want to run bigger models you can look at quantization etc, or just investing more money.

[-]

420GB@reddit

Most companies can afford a couple hundred bucks up front cost for a NPU or GPU

[-]

sofixa11@reddit

NPUs are mostly useless, and GPUs to run large enough to be useful models do not cost a "couple hundred bucks". Nor does the rest of the machine they have to go in.

[-]

Bazzatron@reddit

Well, Claude just inadvertently went open source. The cost of hardware to run it and make it available internally is going to be lower than the cost of a major data breach.

The point being, how do you enforce this when it’s being done on a BYOD device that’s is allowed applications outside the scope of MDM policies.

I mean isn't this no different than the dawn of time of employees who can see payroll, sensitive trade data, insider secrets?

You start at the hire training suite pushed through HR, it is stuffed down your throat (at my bank) repeatedly at the start of any role YOU DO NOT EMAIL YOUR SELF WORK DOCUMENTS. Alongside it being fed to you at the start there is an annual course you have to do touching on this subject, its the same here.

We do it for calling from your personal phone number to clients, texting coworkers company related conversations, etc.

[-]

improbablyatthegame@reddit

It’s definitely within this context but supercharged and the ease of breach is only going to make these instances more dangerous and accessible to end uses.

[-]

cyber_r0nin@reddit

You realize the more devs use these tools the sooner these devs will be out of a job. The whole purpose of using the AI is to train the AI at the same time. Eventually the AI will get well enough at development that a developer may not be needed. Instead of training your human replacement (some new intern) you're training a non-human Ai. Granted this will take a while...

[-]

ride_whenever@reddit

The point this will happen has very little to do with the ai tooling as such.

It’s all about the surfacable content, and being able to give it context. Most companies just don’t have that available, but the AI is relatively good at producing that as it goes.

Once the tools are widely enough adopted, we’ll see companies coming up that have entirely grown with this tooling, and the associated documentation/contextual layers - that will be the point that you’ll see the end of current software dev.

Same with sysadmin work, once companies come up with their infrastructure established with contextual cues, that’ll be the end of it.

However in the meantime, I’m loving being able to work at the speed that I’m thinking, rather than being slowed down by the tedious act of execution. It’s exhilarating, and I’m massively enthused with my job again.

[-]

cyber_r0nin@reddit

I understand that, just...change is annoying...

[-]

Substantial-Fruit447@reddit

My org set up CoPilot as our approved internal AI service.

We can link to our Outlook, OneDrive, feed it company data, within reason.

We were told "nothing confidential" and "no API keys or client secrets, or credentials otherwise."

Obviously it's not going to stop anyone from feeding stuff into another LLM on another device, or taking a picture of something on their personal phone and sending it into Claude.

You could go the route of blocking it, but that's a lot of work.

Instead, this is something to be handled more at the policy and management level.

Clearly define which services you can and cannot use and what data you can and cannot put into it. Anyone that violates that policy will be dealt with administratively.

[-]

ansibleloop@reddit

I love the Copilot comedy clown show in VS Code

The auto-prediction is enabled by default, meaning whatever you type is sent to Copilot

dynalisia2@reddit

I’m curious: are you using Microsoft cloud services?

Superb_Raccoon@reddit

Very large companies use copilot for "individual" use cases. Even IBM does, as for commodity AI it is hard to beat.

[-]

fxs38@reddit (OP)

Thanks for sharing your experience! Very useful

[-]

Biglig@reddit

It’s an ancient anti-pattern. Policy based security controls don’t work if you don’t also have an effective disciplinary process for when someone breaks policy. Fire this guy noisily and it won’t happen again.

[-]

Ballbag94@reddit

Fire this guy noisily and it won’t happen again.

They should have a properly defined policy along with what is acceptable, not just fire the dude to make an example of him otherwise it will happen again, just not the exact same thing

"Don't put company data in AI tools" is way too broad without even defining what is classed as data and what isn't. They need both halves to be there, not just the disciplinary half

[-]

Biglig@reddit

Oh yes, it has to be “we fired him because he didn’t follow our clearly written policy which we’ve made sure you all read”.

[-]

KJatWork@reddit

IT has to get out in front of new tech and drive it or someone else in the company will.

Many years ago, wireless access points were still new and our company's IT was concerned about security, so they refused to deploy them to offices, but they also did nothing to physically secure the IT closets in the offices either...90s were kinda wild. Office Managers would head over to Best Buy and pick one up on the shelf, plug it into the hub in their IT closet and pass the default password around to their office staff and any customers/clients that needed/wanted internet.

You may not like something and it may be a risk, but if you are driving it, you control it. If you don't, someone will and that someone is right now copying company data into public AI tools because you didn't give them a company option supported by policy and leadership to use instead.

[-]

Beneficial-Gift5330@reddit

Given that WiFi didn’t effectively exist until mid to late 1999, I am finding this comment mildly amusing, if inaccurate

[-]

HesSoZazzy@reddit

PCMCIA card in my laptop connected to a Linksys WRT54G. Blew my fucking mind the first time I was able to get a page to load from my chair in the living room. :D

[-]

Beneficial-Gift5330@reddit

Right I don't doubt it happened, just not in the way he described. Both because he's violently wrong about the decade, but also because laptops were pretty rare at that point and installing a PCI card in a desktop was not as easy and clean of a process as he described an office manager doing. IT guy? sure. Office manager? Absolutely not

[-]

gp24249@reddit

this - I remember !!

Vektor0@reddit

I am flabbergasted by the number of wannabe IT professionals in here responding to the most obvious AI post ever. None of these people have any business having IT credentials or anything to do with cybersecurity.

[-]

upq700hp@reddit

seriously...what the fuck. not the first time either

[-]

upq700hp@reddit

I get so fucking pissed everytime I see someone complain about AI using an AI written text. Is it engagement farming? Are they actually just stupid?

So many plausible explanations, not one of them reasonable in the slightest.

[-]

iamLisppy@reddit

— is the dead giveaway. Nobody actually uses this in the real world.

[-]

I wouldn’t say not knowing three different types of dashes is mildly illiterate but sure — I can take my lumps.

[-]

civbat@reddit

Harumph. Wholly ignorant of Dashes. Utterly disgusting...

[-]

stephenph@reddit

I have just found out about them and — started — using them.

[-]

Ur-Best-Friend@reddit

and — started — using them.

Incorrectly—apparently.

[-]

PAXICHEN@reddit

Samsies.

[-]

chriscrowder@reddit

I use semicolons

[-]

8inches_inside_daddy@reddit

Excuse me - I’ve been using hyphens religiously before ChatGPT was around.

I used to use two of these: -, but then text editors started smooshing them into a big long one. And in some places it will look right until you send.

rainer_d@reddit

Our policy: „Don’t ask, don’t tell.“

All our mails are in Exchange Online, the Clients have Falcon agent installed, most servers too.

If management was worried about any secrets spilling out, they would have to stop doing a lot of shit before they can worry about Anthropic , OpenAi and all the other companies.

[-]

gbell76@reddit

If you are running an M365 environment, you may want to consider co-pilot licensing with OpenAi or Anthropic as a subprocessor. This way GDPR/HIPAA/ITAR/etc compliance will flow through considering that you have a BAA with Microsoft. It will also keep it logically siloed and you can somewhat manage it. Just a way to CYA and at least give the air of you doing SOMETHING to maintain compliance and best effort. The alternative in the Wild West.

[-]

[-]

Fatality@reddit

Why are you not giving people the tools they need?

[-]

bezerker03@reddit

This is a policy issue. Start officially writing people up and or firing for it. They are basically sharing company code or private info with the internet.

[-]

Total-Assumption-494@reddit

Data loss prevention to monitor endpoints for uploads to cloud services

[-]

Ok-Reply-8447@reddit

Managers should take the time to raise awareness and provide proper training so everyone understands the risks and consequences of sharing sensitive or critical information with GPT tools.

At the same time, policies need to be updated to reflect this clearly.

lotekjunky@reddit

zscaler ai/ml classification is fast enough.

[-]

cometwrench@reddit

I mean, if you’re just giving them standard copilot then you may as well block ai outright since its an active hindrance as opposed to tangentially useful every once in a while.

[-]

matjam@reddit

Get agreements for the tools and provide access through the self service portal.

Most of them are able to do a “we won’t use your data” agreement.

They probably lie of course but it makes the lawyers happy.

[-]

DopamineSavant@reddit

like it was nothing.

Well you all don't have a policy, so it was nothing.

[-]

cyber_r0nin@reddit

OP stated there is a CYA 'line' in the manual - that no one reads.

Not reading the company handbook isn't an excuse. And most businesses require you to read the manual for this very reason. If there is an update then an all hands e-mail should be sent out indicating there has been a change. If that isn't enough then require EVERYONE to read the manual as a yearly training. If they sign off that they read it (whether they did or not ) the business is covered.

Orgs need to stop giving passes to this nonsense. There are *plenty* of workers in the waiting for a job. If you use fear in this instance it might work just enough. But you have to be willing to back it up. Someone breaking the AI rule will have to be made an example at some point. Otherwise there is no teeth to the no AI rule. In this instance since the fix is in some sort of measure needs to be taken, probably not a complete firing. (emotional decision - maybe they have a family(?))

[-]

SandmanPC@reddit

I implemented cloud access security broker (CASB) policies with Data Loss Prevention (DLP) using Secure Access Service Edge (SASE) product called Netskope.

We blocked requests that contained specific information, keywords, health or personally identifiable information. It allowed fingerprinting data and matching that for DLP rules.

Powerful, yet expensive, product.

[-]

lotekjunky@reddit

zscaler too

[-]

fxs38@reddit (OP)

That’s a heavy setup ! But yes it seem to do the job

[-]

junon@reddit

Most secure web gateways offer this functionality now. We had it with Cisco Umbrella before and we have it with Zscaler now.

[-]

SandmanPC@reddit

I find Netskope much more intuitive then Zscaler. Ive heard rave reviews. But when i finally got my hands on it, i didnt enjoy the experience as much as i did with Netskope.

[-]

gumbrilla@reddit

What the hell are your management doing???

Go and buy Anthropic Enterprise
Give claude to developers.
Usual users, you could give them copilot, but it's shit. Probably give the Claude to them also tbh. that's what we do as well as copilot, but it's shit. shit shit shit.
Put in a policy and technical controls (purview, network monitoring)

You have unmet demand in your organization, if you don't meet that demand.. like water, it will find a way.

[-]

cyber_r0nin@reddit

Unmet demand?

First, what would have happened if there was no AI?

Second, why is their well paid dev not able to figure out the issue?

Third, if 1 guy can't debug why isn't there a team or a backup dev also able to help?

AI IS NOT THE FIX.

ImpossibleLoss1148@reddit

Places with a coherent security strategy are leveraging vendor tools to scan for rogue AI in the enterprise. Places that just fall into the next crisis, are not.

[-]

nmsguru@reddit

We have created an AI policy document a year ago with specific tools allowed. Employees must sign it before getting access to AI. All the rest are blocked by FW and proxy. Employees that want to use AI need to submit a formal request, approved by their manager - get a licensed Enterprise account for the AI tool with their corporate creds. Additional level of DLP is enforced through a dedicated software that scans AI prompts. Cyber team reviews violations and alerts the managers. Employees that paste PCI and other secret stuff are called and warned. No layoffs yet but they behave after a few got caught and warned,

[-]

A couple of things we're thinking about at my tiny company:

- MDM policies to only allow company-approved AI tooling, which is basically Gemini via Google Workspace, and Claude Code/Cowork for engineering.
- We're thinking about the guardrails and sandboxing options in Claude, and how we can roll out recommended security config.
- Tracking which MCP's are enabled across which SaaS platforms.

These were recommendations made to our IT team, as I am more of an engineering security guy, not a sysadmin.

This GenAI era is just highlighting how much of an issue shadow IT is to not just manage, but also have visibility of. We really need an Eye of Sauron on all of it. Not sure if Vanta/Drata etc have made any progress in these areas, but I guess we're always limited by what these AI/SaaS providers allow you to check from their API's.

Bonne chance, my brothers and sisters in tech.

[-]

gogetit57@reddit

We use Lightspeed filtering (education). They are bringing in an AI Monitoring tool that gives exactly what you state - which users are using which AI platform and what prompts are they entering. Whilst you may not have the same filtering requirements, this module may help you in your quest.

[-]

TommyTheOneAndOnly@reddit

We have Copilot as the approved AI platform, chatgpt and some other tools are blocked so the site isn’t reachable, works for the majority of

[-]

Popal24@reddit

You have to give some leeway and provide corporate AI dev tools. It can be corporate Microsoft Copilot or something else. This will destroy the need to use private tools.

This, plus education and policies on why it's bad and with appropriate sanctions if it's breached.

I have access to enterprise licenses options the company has negotiated or internal options. Everything else is not allowed

[-]

critler_17@reddit

Best answer for sure. I have my own chatbot my employees use that’s just a ChatGPT skin and will chew them out and let me know if they’ve done something naughty (not really it’s just like hey don’t do that and I get an email)

[-]

GeoSystemsDeveloper@reddit

Block it from their laptops ...

[-]

Benvolix@reddit

The most effective solution here isn't surveillance, it's substitution.

People will always reach for the best tool available when they're under pressure at 11pm. If you don't give them a managed alternative, they'll use their personal accounts. Block it on the company device, they'll use their phone. It's not malice, it's just human nature.

The fix is to make the managed option the path of least resistance. Get them GitHub Copilot, ChatGPT Enterprise, Claude for Enterprise, or whatever fits your stack. These give you the data protection guarantees you need (no training on your data, tenant isolation, audit logs) and your devs get the tool they already want to use anyway.

Once a sanctioned option exists: - You have actual audit trails - Data doesn't leave your controlled environment - You can enforce acceptable use through the platform itself - Employees stop hiding it because there's nothing to hide

The dev in your story wasn't being reckless - he was trying to do his job. He just had no good option available. That's a tooling gap, not a culture problem.

Policy alone never solved shadow IT. You need to close the gap that creates the incentive in the first place.

[-]

Inn0centSinner@reddit

Management folded like a jellyfish, "please unblock the AI URL category".

It's not my job to take on risk, it's my job to communicate proper security posture and it's management's job to decide if that's what we're going to do. I'm sure some of our PI is floating around in several LLMs by now.

[-]

dllhell79@reddit

I have seen 2 promising solutions for the general problem. One is Palo Alto Prisma. It's a managed custom Chrome browser that lets you directly apply firewall policy and dlp rules to traffic. Since it's in a browser, it's all encrypted by default, so visibility is pretty much everything. You can do things like prevent flagged sensitive data from being pasted into an AI engine, etc. A very nice solution... but with a very nice price tag.

The other is Prompt Security. They were purchased by SentinelOne last year and are being integrated into the Sentinel product. I believe it's going to do a similar type of blocking to the Prisma blocking, but just at the AV level instead of directly in the browser.

SirLoremIpsum@reddit

Not looking for "block ChatGPT at the firewall" answers — we've been down that road, it just makes people use their phones or hotspots. I mean actually tracking and managing it.

You can't.

You literally can't.

If someone has a phone you can't stop them from typing it in right?

You can only block, restrict and deal with it at a policy level "any usage will be subject to disciplinary action".

If someone can see the code or the data, they can type it in to chat gpt on their personal device.

[-]

EngineersOfAscension@reddit

Monitor who visits chatgpt personal via firewall/web filtering rules.

Raise it with HR and/or their manager

[-]

wtjones@reddit

Everyone needs access to a safe LLM for them to put whatever they want into. This is the dumbest shit in the world. I work for a huge backward company with 12,000,000 stupid security policies meant to keep anyone from getting any work done and everyone has access to a safe LLM they can put anything into. This is like telling someone they can’t use a computer or a cellphone at work. These are the tools of modern work.

[-]

og-golfknar@reddit

I believe it's best to communicate why it doesn't work for your company first. Get their buy in. As you build a company based AI which can be tracked.

But I assure you don't have the power to make anything like that happen nor it feels like the ability to conceive next steps without direction. But I hope I'm wrong and you take my advice.

[-]

PrestigiousShift134@reddit

ChatGPT ships with MDM controls

[-]

We have approved AI programs that such as ChatGPT and Copilot that don't train or share company information to the public model. Our DLP solution blocks company information from being shared on other AI programs. If they try to share company information on any AI model it will be blocked.

[-]

Loop_Within_A_Loop@reddit

we have a ChatGPT Enterprise subscription that any user can install and use, and rely on OpenAI's data governance to protect our classified data

banning a program as easy to access as ChatGPT for security reasons is just asking for users to find a way to circumvent it, make your execs pay for it

[-]

qzjul@reddit

This is the way. Give the people the tools they want, but the safe way.

People keep leaving for half an hour for coffee because the nearest coffee shop is a fifteen minute walk away? Make coffee for them in the office...

Some "perks" cost the company less than the alternative.

[-]

Thirsty_Comment88@reddit

Just stop giving a fuck.

It would be nice if the AI vendors would provide vanity domains for business accounts so that we could block everything except that URL. But, like you said, if you block it, people will just use their phones or some other device.

The only thing you can realistically do is provide company AI subscriptions to employees that provide the tools the employees want to use. And immediately fire people who use personal accounts an get caught. It has to be a zero tolerance policy, and you might even lose a good person or two. But most people will choose the path of least resistance, which is to log into their company provided AI tools, not their personal free accounts.

[-]

YungMotch@reddit

Look into ZScaler. You can configure data loss prevention policies that will detect if anything sensitive is being sent over any AI/ML tools

[-]

AndyceeIT@reddit

It's a policy issue, don't try to solve it like a technical problem.

Raise it with management. If you can't convince them to care then ask for that in an email & stop caring
Get a standalone "Acceptable use of AI" policy written. Plenty online you can copy, or get chatgpt to help 😉 . Get contributions & buy-in from a range of supervisors. Be clear about what is specifically not acceptable, and
Presuming you were successful in getting management on-board, ask them to disseminate it, or at least formally endorse.

fin

[-]

FoxtrotOscarBravo@reddit

They said, if you can’t fight it, join it. My organization has subscription to ChatGPT enterprise. They highly encourage people to request a license and join the enterprise workspace

[-]

CtrlAltDust@reddit

Block everything on enterprise endpoints and use HIDS to monitor data exhilaration.

Also if you can lnk your chatgpt enterprise tenant to zscaler, and do SSL inspection. Allow only access to the chatgpt workspace with your enterprise chatgpt tenant while blocking chatgpt cloud app.

Put the allow for your chatgpt tenant above the app block policy and it works wonders. Can assign authorized chatgpt security group to control exactly who can use it.

[-]

Acceptable_Mood_7590@reddit

You can block coppy but allow paste which is useful and won’t cripple your staff

[-]

Electronic-Jury-3579@reddit

Get the enterprise plan for one or more of the chat gpts and trust the encryption they offer. Disallow other uses.

[-]

apriliarider@reddit

1 - policy and enforcement

2 - communication and training for said policy

3 - ai proxy w/query DLP capabilities, and AI access policy limits.

4 - block all other AI

Sorry for being short. On Mobile.

[-]

BankingAnon@reddit

We use an app that strips sensitive data before it hits the LLM. Prior to implementing this, it was very limited on who even got access, we also have a CASB in place to cover accidental leakage / DLP stuff.

[-]

pantherghast@reddit

Use co-pilot which is managable via Intune and auditable via Microsoft purview. Block all other AI

[-]

idknemoar@reddit

Microsoft Purview DSPM for AI, blocking any sensitive data from being placed in the various ones. May not be perfect, but it’s included in our E5 licensing.

[-]

MReprogle@reddit

Literally the only thing I know of right now, but like you said, it’s still easy to get around. Best approach if you have the money would be a cloud gateway or internet proxy and block everything going to the domains of your choosing, or redirecting them automatically to an approved tool.

The bad part is that Purview takes quite a bit of time to properly roll out without breaking everything, and needs to be reviewed by more than just IT. It’s not like one DLP expert is going to know everything about every bit of data to even know what to see classifications to, and auto labeling data would likely end up wrecking life if they tried. I honestly don’t think most orgs know how complex purview can get .

The scary reality is that most companies aren't handling this at all and don't have the slightest concept of how to even start handling it.

The best solutions I have seen though is simply blocking the LLMs at the network level and only allowing company approved ones. For instance we allow copilot since it comes with our 365 licenses so that is allowed but you have to be logged in with your 365 account. We have forced redirects in place for copilot that take you to the corporate login page so you have to use your employee account to use it.

Will I ever use this? No. Should companies consider this rather than blocking the domain entirely? Some should yea.

[-]

mrfoxman@reddit

Look into hatz or the probably dozens of other services doing something similar. Where they supposedly don’t train on your data. But take that for what you will.

[-]

Thrashtah_Blastah@reddit

Cisco Umbrella was a good starting point for us. We have the agent installed on all company workstations with a global policy blocking all AI. Copilot is the only exception globally (personal site blocked still). Departments wanting access to unapproved AI sites must submit a request. After review it must have signoff from their management and IT management. Then it gets added to the appropriate allow policy to supercede blocks.

That put a dent in the problem for us. There is always going to be an issue with folks skirting around it with personal phones. That's a policy violation though and becomes an HR issue.

[-]

snottyz@reddit

Policy is that staff can only use enterprise accounts. When we blocked access to other tools (including ChatGPT), the revolt was IMMEDIATE and FIERCE, so it was allowed again. Soooooo not handling it well, overall.

[-]

Crowdh1985@reddit

You can enroll ChatGPT in Intune… and buy licences. Otherwise Copilot with licences can use ChatGPT in settings. I do have both and even if Copilot can work with GPT, GPT is way snappier then Copilot (cloud or local)…

With intune/purview you can manage how data is managed and especially office apps. No more copy/paste on personal devices , which will force them to use office devices. No more GPT, well intune will block it… no need to go in firewalls!

Intune and purview is the key!

[-]

audixe@reddit

Zscaler can apparently be used to detect this and block it. My company uses Zscaler and I can see AI sites being used but we have no formal policy blocking it.

If this is a concern to you it’s definitely worth looking in to.

[-]

Creative-Type9411@reddit

there is a certain point where you make something policy and if you discover they break policy, they get the axe

It is exhausting trying to child proof everything

[-]

kiddj1@reddit

We have a subscription for GitHub copilot and we access that through a custom URL

Every other Ai domain is blocked full stop

[-]

mixduptransistor@reddit

There will be a lot of technical answers and solutions that eventually hit the market but a stern “do this and you’re fired” policy will go a long way. At some point you can’t prevent someone from picking up a personal device and doing something they shouldn’t and that’s where HR takes over

[-]

fxs38@reddit (OP)

Good point !

[-]

sobeitharry@reddit

Yeah, first you need a policy in writing. Then, if leadership approves the budget you implement any of the multitude of potential technical controls and tools available.

If someone intentionally violates policy and/or controls, they are written up or fired depending on the level of the offense.

Blocked all ai sites based on classification. Then only allowed 365 Copilot. With enterprise licenses the data is kept in your tenant.

You can then use Purview to do auditing on it.

[-]

tfn105@reddit

This is exactly what we've done too. Completely achievable

[-]

OneSeaworthiness7768@reddit

Are people running anything to get visibility into which AI tools employees are using?

This post feels a lot like research for someone who’s working on a tool to do exactly that.

But the answer is no. My firm gives employees enterprise plans to use so they don’t feel the need to upload company data to personal accounts to do their work. Not sure why a company would pay for something to give visibility into what employees are using on their own but not pay for the tools they actually need.

[-]

Superb_Raccoon@reddit

Fire the fucker if you had a clear policy in place.

If you didn't, make it clear it is a firing offense and firemen next fucker to abuse the trust of the company.

We've been digging into this quite a bit and the the front end it's policy. You can basically divide that into 3 areas:

Generative (what you mentioned)
Agentic
Note takers

We separated note takers out because there's a slew of things related to them (consent, retention) that are different than the other areas.

I think the key thing right now is to approach this from a policy perspective and lay out guidelines on the use of those items. As far as being able to automatically stop someone from using it, the security tools just haven't caught up.

[-]

fxs38@reddit (OP)

Interesting share, thanks. Policy is key, education of users too. Haven’t seen a solution to stop, perhaps alerting would be interesting as a start - reminders to users when they visit such sites

[-]

Bourbonneuxb@reddit

The way my company deals with this by using copilot then blocking all other AI/gen AI sites and have policy exceptions for specific use cases if copilot dose not work for whatever the requirement out of a few thousand employees there are 4 exceptions.