From context, it seems there was an API that was internal for support use but was supposed to be gated by some required process of convincing the support agent you were who you said you were (also vulnerable to social engineering) but they didn’t really evaluate whether tools intended for conscientious human use should be provided directly to the LLM that replaced the former support agents.
It's insane the AI has been provided the tooling to send emails to arbitrary addresses like that. Like, getting it to send a 2FA code at a user's request is one thing. But it should only be able to "hit a button" to send a 2FA email to the address attached to the account, all run with hand-written code. It shouldn't have access to the 2FA code itself, or the message subject, or body, or the recipient address, etc.
This is not true. Well, it kinda is, but nobody will be stupid enough to hand-code an account recovery where you get to type any email address.
The reason it worked there is that the designers of the system didn't anticipate that the AI will agree to accept any email (maybe they even put guardrails against it in the system prompt, we don't know). It's more like social engineering than bad-security-code, except that like the sibling comment said an actual human will probably not approve that.
My impression is that AI didn't replace static code in this place; it replaced a person, who (hopefully) would have been suspicious about sending an account recovery code for e.g. "obamawhitehouse" to e.g. "bscurtu.alfamm.ro@gmail.com"
I do a lot of bug bounty research on Meta and Instagram, and some of the bugs I find look extremely simple like this but have some slightly complicated reason for why they occur. Maybe not this one, but I do have a guess as to what might have actually happened.
Based on what I've seen so far, Meta AI Support Assistant (they call it "MAISA") had tool calls that a) start an email verification to any specific email, phone number, or the contact points linked to an account and b) allow generating a password reset link for an account based on an email verification attempt. I don't think it had any access to the actual codes themselves, but rather think a handle or ID for an email verification attempt (along with the user provided verification code based on user input) was provided to the "generate reset password link" tool call, and the tool call failed to properly validate the actual email used in that attempt belonged to the account allowing the ATO.
The tool call for MAISA to generate a password reset link should have failed with an email verification attempt that corresponds to an email not linked to the account (and I believe I even tested this at one point on Facebook and encountered an error that successfully prevented it), but I suspect they tried making a change to this tool call for Instagram where slightly older, recently unlinked emails could be used to recover an account that got hijacked by an attacker, which added the need to allow emails not currently linked to the account to be used and set to the user's primary email.
I also suspect that the MAISA tool call change called a wrong API or something that unintentionally allowed any email verification attempt that was successful to be used, but the engineers did not add a sufficiently thorough e2e test case to test the tool call against unrelated email verification attempts being provided to the tool call. This is the part I think should be focused on the most. Tool calls for agents that have their output potentially influenced by an attacker should be treated like external APIs that anyone can reach, and they should be tested as such.
This is all obviously a guess, doesn't take into account the many signals they use to determine if an account recovery attempt is valid, and could be very inaccurate, but it's the closest to what I (someone who deals with Meta security a lot) think could have allowed this to happen.
Some Jr engineer got tired of handling stupid support requests and automated the job with an agent. That’s how.
Assigning Jr engineers for security support is ridiculous partly because young people don’t understand how critical security is sometimes. And partly because they don’t value privacy as much.
> But it should only be able to "hit a button" to send a 2FA email to the address attached to the account, all run with hand-written code.
Genuine question...why would that need to be hand-written?
It makes absolute sense as a general statement and is kinda crazy that this wasn't a built-in limitation, but I'm not quite sure why the code for that bit must be hand-written (provided the code functionally does what you describe).
I think he likely means "code that is hand-reviewed" and not directly controlled by the agent. He's probably meaning to differentiate it against the in-process agent writing the code. It doesn't matter too much if that fixed code was written by an LLM under guidance and review of the SWE, outside the agent.
This exploit is my new gold standard for trivially avoidable security failures. Someone has finally beaten Gitlab's password reset emails to attacker-provided addresses.
I'm among the first 6000 users of Instagram and my first name username was stolen a few years ago. Support for verified accounts acknowledged the issue, but couldn't do anything about it.
This turn was an AI exploit, in my case was an outsourcing support 'exploit', where someone paid for my username to be manually changed and given to another user. There will always be a way to get access to accounts if human accountable support doesn't exist, with criminal consequences for employees that violate it.
Support requests have always been the weakest link in the security chain for big corps. I've had accounts of mine turned over with 2FA disabled by humans before. I guess we shouldn't be surprised that the LLMs are doing the same thing.
The simple fact that 2FA can be removed by low level support staff drives me mad. It defeats the whole purpose of the process.
>> The simple fact that 2FA can be removed by low level support staff drives me mad. It defeats the whole purpose of the process.
The fact it can be removed by anyone is the problem. If you lose access to your 2FA (and recovery codes) then you should lose access to your account. Having it removable by anyone (other than a logged in account holder) defeats the entire point.
It's all there, and high-stakes environments with no proper protocol are most vulnerable.
Source: used to work part-time in IT support at a hospital, by now 10+ years ago, so it was routinely requested to circumvent regulations and security protocols, even medical ones (cough Windows in ICU monitors and other medical "kiosk" PCs that should absolutely not run Windows)
This is not wrong but what’s really missing is cost: Meta did this so they can avoid paying people to do it. Lots of companies follow that decay spiral: your bank could shut phishers down cold by requiring wire transfers to be authorized in person but they don’t want to pay staff or risk you being upset by a transaction taking an extra hour so they don’t.
Imagine an alternate universe where big tech companies worked with various trustworthy third-parties where something like this would generate a challenge you could take to your local notary, post office, library, police station, etc. where someone would check ID before approving it. How many phishing attacks would be prevented annually by a physical presence check?
> your bank could shut phishers down cold by requiring wire transfers to be authorized in person but they don’t want to pay staff or risk you being upset by a transaction taking an extra hour so they don’t.
Isn't this essentially what just recently happened to the Pope? Then there were people here doing the rest of your comment for him saying how egregious it was for them to ask for an in person authorization. It sounded like all he was trying to do was update his address, but changing your address from one in Chicago to one in a European country absolutely sounds like something a phisher would be trying to do.
Then you get trusted parties selling account access. Even if you remove them for a single false positive they will do it. A bit like a % packages "vanishing".
for a while facebook had the ability to recover your account by having them ask several of your friends if the recovery was legitimate but it was turned off. my guess is that not enough people added trusted contacts to bother running it.
The amount of hassle involved with regular physical checks is why it's not implemented, regardless of attack prevention.
The cost of hiring a person is part of it but not really the core reason. People were sold on the Internet with "you can do things online conveniently" and reintroducing the need to physically go somewhere negates that angle entirely.
> People were sold on the Internet with "you can do things online conveniently" and reintroducing the need to physically go somewhere negates that angle entirely
But how often does one need to do recovery procedures like this?
How much less convenient is it for everyone else to be at risk of their account being taken over?
It's a hard problem. How do you prove you own an account if you lost all proof of ownership? Especially so if an account was never tied to your real name, in which case you could at least rely on government ids.
I had to go through the account recovery on my Facebook account once and the proof they demanded was that I match a bunch of pictures of friends to their names. I think it took 3 tries over multiple days to actually get it unlocked because it turns out I such really remember a lot of the people I met 20 years ago and friended on Facebook.
I don’t recall why I had to go through this song and dance. Very plausibly the account was still associated with an old school address that I could no longer access. So yeah, account recovery is hard. How do you prove someone owns an account when they’ve lost the things they are supposed to use to prove ownership?
I manage customer identity and access management ("CIAM") for a financial services firm. Passkeys are primary, recovery can be performed by providing a government credential remotely (which costs us ~$2-3 per recovery). I do not think it is hard, based on what we have built and spent to enable these capabilities. NIST Special Publication NIST SP 800-63 Digital Identity Guidelines is a helpful resource on this topic.
I think Meta just does not care if they're enabling AI attack surface and vulnerabilities into these customer journeys. It's...certainly a choice, versus deterministic journeys with hard guardrails. They could make different choices.
> recovery can be performed by providing a government credential remotely
That only works because you presumably do KYC when you open accounts, so you have an identity to match to. Most internet accounts don't do real KYC, so a government credential doesn't really work for recovery --- they didn't know who you were, so proving who you are doesn't help anything.
That doesn't mean that letting anyone sweet talk support or an AI into taking over an account is acceptable, of course.
It's a fair point, and can be solved for as part of the "Verified" offerings Meta offers. This binds IRL identity to the digital identity at verification for future identity assurance step up (including if and when recovery is required). Failing that, TOTP, SMS, and even mailing an OTP to a mailing address remain low friction auth factors (with, of course, various levels of security).
My point is that while this is not easy, there are obvious very bad ways to implement this that should not be done (chatbot or other generative AI interface vulnerable to the usual suspects of AI inherent attack surface). Don't build the bad way, the right away is known and straightforward.
Range != value, depending on use case. Doing more poorly does not make something better. Our customer identity capabilities are very close to login.gov (we don't have to support hundreds of agency customers and common access cards), and if its good enough for ~342M Americans, its good enough for our customer base.
Broadly speaking, work for the sake of work is not valuable work. Show me outcomes for resources and time invested, and compare accordingly. Value is, again broadly speaking (there is always nuance), what you deliver. If you bring me an AI solution for a high risk high value customer journey, data flow, or code path, that is an anti pattern. If you, as a colleague or a stakeholder, put forth that we must use AI in situations that require a high degree of determinism (due to potential high cost failure modes), you will need to prove this extraordinary claim with evidence.
Choose Boring Technology - https://news.ycombinator.com/item?id=9291215 - March 2015 (212 comments) ["Am I using this project as an excuse to learn some new technology, or am I trying to solve a problem?"]
I get paid to manage risk efficiently, including being measured on time and budget spent against the success criteria, ymmv; my comp and budget is not dependent on how much AI I shove into security systems. "What am I optimizing for?"
the alternative is people losing their accounts and people aren't willing to allow that. i do think that apple does this a little better where they try everything to contact you in every way they know and it takes a week to get access. at a minimum to change your email it should require a week of waiting to see if the user can access the original mail to the hand off.
So the AI agent had privileged access to remove 2FA, ignore the account email, and just hands accounts to whoever asked? Honestly that’s so highly negligent I wonder if the implementation team for that “feature” was intentionally trying to do as much subtle damage to meta as possible before their inventible layoff.
It’s a shame nobody tried to get it to drop the production table entirely! (mostly joking). Just claim to be a high level SRE solving some critical production bug, the only solution to which is dropping the database.
I get that account recovery for sites with hundreds of millions of users is a huge burden they're struggling to manage but I'm shocked they didn't restrict such loose verification to the >90% of lower value accounts that aren't worth stealing and keep the stricter verif on high-value accounts.
The next obvious thing would be to let accounts the algorithm judges to be low-value still opt-in to strict verif. The vast majority of low-value accts won't bother flipping it on if the option is buried two menus deep, but many of the few low follower/views accts who are targets for some other reason (political, stalker, etc) - know they are targets and can self-protect by opting in, further reducing account hijacks.
So, before we even get to whether this 'loose' verif is "bad", those two simple implementation changes would certainly have cut the bad outcomes of a (potentially) bad idea by >95%.
This is true for any service that Meta owns. I experienced something similar on my Meta (formerly Oculus) account. Meta support is very susceptible to social engineering and they have been for some time.
Always a bit illuminating to me how many exploits seem to so dumb I'd never even bother to attempt them. You're telling me I can just...ask for the password? And that works?
This is an embarrassing failure for Instagram. But SIM cards have been hacked the same (by tricking support, claiming the phone was lost or stolen), except the agent was human.
The solution (which also solved SIM support agents being bribed or hacking known acquaintances) was to prevent the agents from resetting the SIM card without some steps the original owner would have to follow (and could follow even if they've lost their original phone), like a PIN they'd have to remember. I think the same solution should be applied to AI agents.
>Once it looks like the request is coming from the correct region, they tell the Meta support AI that the account is hacked and ask it to send the verification codes to an arbitrary email address they control.
Dear Instagram, wtf. Why not send the reset to the account in question? Arbitrary email, wow.
Perhaps the attacker says that they email was also hacked and "this is my new email now". It sounds like this was a result of AI support and not a real person "And if you're part of the A/B tested accounts on which the AI support option is active, tough luck, you can't even turn it off."
This happened to my instagram yesterday night while I was asleep. I don't have a particularly high value username (it's probably worth somewhere in between $300-500), but still incredibly frustrating to deal with. True to the article, I had already enabled 2FA last night and it didn't matter.
Thankfully, IG gave me the option of restoring my username when I logged back into my account today.
The ironic thing is I know several legitimate humans who have lost access to their accounts and have been trying for months/years to get restored access, but being stuck in support hell.
I'm sitting here wondering why the Chief Master Sergeant of the U.S. Space Force has an Instagram account to begin with. I understand it's the office itself, but still don't see the reason to expand the attack surface of government offices. X makes sense, Instagram, I'm not so sure as much
Outreach, I'd guess? You've got to do outreach where the people are. X and Instagram have pretty different audiences, but they're both large, so if you're on one you probably should be on both.
I see no difference between X and Instagram in this regard whatsoever.
Think NASA, for example; it's also a government agency, and they are doing great job posting photos in Instagram, do you think anything is wrong with it?
Nothing says you are an advanced stupid company than using AI to implement the stupid. This is security I doubt even a college student would implement. Does Meta have a CSO? The correct answer is they don't, even though some body might occupy the title.
Of course it's always possible that they simply don't care who has your account, as long as they get money.
Just another day for Meta in terms of embarrassing outcomes, and yet the company makes hundreds of billions of dollars per year because the only thing that matters anymore is shoving increasingly scammy and worthless ads in front of as many eyeballs as possible, even when the people with those eyeballs can less and less afford to buy anything non-essential.
I suppose you could chalk this up to an oversight. I don't see how Meta gained from this. They've been purposeful about collecting user data and lying about it, eg: 2025 Android Tracking Incident. Shouldn't just be an embarrassment, should be much worse than that.
What is even the point of having 2FA if it can be so trivially bypassed? Isn't that the whole point that it's sort of a last line of defense? Oftentimes, you can't change simple account settings without having to re-auth and then punch in your code again. Why would something as critical as a suspicious password reset be able to jump ahead of that? Mind boggling. But, I guess that's what happens when you lay off 10% of your people at a time.
If the LLM has knowledge of something, by design it can't help but divulge it. When will companies learn granting any kind of sensitive information access to an LLM is a moot point
What part of this article implied the LLM divulged sensitive information to a user? All it did was change your associated email if you impersonated the user
I think the related news of Meta rolling out subscription models for their free products, is a step in the right direction.
Otherwise the only way to provide these services is to massively underfund support, if you charge 0$ per account and serve 1 Billion users, then you cannot afford to spend 1 minute of human support time on an account.
Yes, they could use the money from ads, but let's be frank, the customers in that case are the sponsors, if the customer is the actual user, then it's way easier to provide direct support to them without facing an foundational incentive misalignment.
My account, with a 3-letter username worth $$$, got hacked yesterday morning probably by this flow, but I did manage to defend it. I think by far the biggest problem with Instagram/FB/Meta auth flow is that 2FA does nothing. You don't need the 2nd factor to disable it, so attackers can just turn it off. Really stupid!
Also, I discovered that many of IG's auth endpoints are just broken. For example you can't change password on web because of CORS, which isn't a transient outage but just a flat out bug.
Edited to add: This is just the cherry on top of years of stupid auth flow at IG. I have received tens of thousands of reset links or codes from IG over the years. There used to be a way to put your account on recovery cooldown for a few weeks but they got rid of even that.
I'm horrified with how poor Meta's use of AI is recently. Here's a list of the issues both me and my wife have been plagued with over the past few weeks. It's really quite an achievement to be this terrible.
1. My personal Facebook received 3 violations restricting my ability to manage ANY Page until April 2027 (lol). The trigger... I deleted 3 unused Pages. These Pages I had created years ago in preparation for projects that never came to fruition, and had never posted any content. THe pages were 'scheduled for deletion', and when that day came (around a month later?), boom, I'm hit with a 1 month restriction which later converted itself into a 1 year restriction after I waited out the month. No Appeal button. I'm expected to wait for a year to manage my new page? All over something that is NOT a violation, just for deleting old pages. Get out of here. Smart system.
2. I pay for Meta Verified on Instagram and for the past 2 weeks "Enhanced support" leads me to a broken interface. "Page isn't available right now". So, what am I paying for exactly?
3. It seems you can use Meta's AI Assistant to sometimes get through to a human. I've done this twice now, and both times my case has been escalated to a different team (apparently) yet I never get an email, I never get an update in the chat (the chat ENDS immediately after the phone call with support), and the issue is never resolved. It's been 2 weeks. The case says "Completed", with no response. Worthless as always.
4. My wife creates content on Instagram and has had her account suspended multiple times now for "Account Integrity". I assume the system thinks she's not the person in the content, despite providing her valid email, phone number, video selfie, and 2 types of ID (passport & driver's license) multiple times. What's hilarious is the passport was accepted on of her accounts (they wiped out everything on her Account Center), but another account was rejected. Great AI, same passport, exact same lighting... different outcome.
So as it stands, we're both fucked on both facebook and instagram thanks to awful AI moderation, and fucked further thanks to awful AI support. No resolution in sight. The incompetence is next level. I really don't see this getting resolved. This already happened to my wife earlier in February, she managed to get one account back, and a month later she's hit with the same identity issues.
Using AI for both the moderation and the support makes me sick. The same poor AI that incorrectly flagged me and my wife's accounts for a load of incorrect bullshit is the same system that's meant to help resolve it? Of course it's going to side with its own poor decision. YouTube seems to do the same thing and auto-reject appeals in seconds. Really smart /s
I believe we need enforcement that social platforms should NOT be using AI to perform destructive actions without human intervention. Noone should ever lose their accounts because of AI mistakes. AI should be used to surface potential issues which get passed to a HUMAN to double check before applying the action. AI simply isn't good enough to have full control.
Fucking pissed off and even angier now I've had to write all this up and remind myself just how ridiculous the situation is. Sorry for the rant, but losing your accounts you put work into is very crushing and demotivating. Being accused of these violations fills us both with so much resent for the companies running this shit.
Sam Cofounder Postmates
On the off-chance there's anyone at Meta seeing this (@Wirah on twitter)
Had to make this new username as my original (samstr) comment doesn't show up. No idea why. Probably shit AI
It sounds really insane. Too bad there is 0 proof or anything in the article, so I am very skeptical. Without proof etc this is just a very nice doom story.
From context, it seems there was an API that was internal for support use but was supposed to be gated by some required process of convincing the support agent you were who you said you were (also vulnerable to social engineering) but they didn’t really evaluate whether tools intended for conscientious human use should be provided directly to the LLM that replaced the former support agents.
It's insane the AI has been provided the tooling to send emails to arbitrary addresses like that. Like, getting it to send a 2FA code at a user's request is one thing. But it should only be able to "hit a button" to send a 2FA email to the address attached to the account, all run with hand-written code. It shouldn't have access to the 2FA code itself, or the message subject, or body, or the recipient address, etc.
Why did they give it any of that?!
This exploit has essentially nothing to do with AI and everything to do with a terribly designed account recovery flow.
This exact same flow could have been (and may have been; I don’t know how much the chatbot here actually does) statically coded.
This is not true. Well, it kinda is, but nobody will be stupid enough to hand-code an account recovery where you get to type any email address.
The reason it worked there is that the designers of the system didn't anticipate that the AI will agree to accept any email (maybe they even put guardrails against it in the system prompt, we don't know). It's more like social engineering than bad-security-code, except that like the sibling comment said an actual human will probably not approve that.
My impression is that AI didn't replace static code in this place; it replaced a person, who (hopefully) would have been suspicious about sending an account recovery code for e.g. "obamawhitehouse" to e.g. "bscurtu.alfamm.ro@gmail.com"
I do a lot of bug bounty research on Meta and Instagram, and some of the bugs I find look extremely simple like this but have some slightly complicated reason for why they occur. Maybe not this one, but I do have a guess as to what might have actually happened.
Based on what I've seen so far, Meta AI Support Assistant (they call it "MAISA") had tool calls that a) start an email verification to any specific email, phone number, or the contact points linked to an account and b) allow generating a password reset link for an account based on an email verification attempt. I don't think it had any access to the actual codes themselves, but rather think a handle or ID for an email verification attempt (along with the user provided verification code based on user input) was provided to the "generate reset password link" tool call, and the tool call failed to properly validate the actual email used in that attempt belonged to the account allowing the ATO.
The tool call for MAISA to generate a password reset link should have failed with an email verification attempt that corresponds to an email not linked to the account (and I believe I even tested this at one point on Facebook and encountered an error that successfully prevented it), but I suspect they tried making a change to this tool call for Instagram where slightly older, recently unlinked emails could be used to recover an account that got hijacked by an attacker, which added the need to allow emails not currently linked to the account to be used and set to the user's primary email.
I also suspect that the MAISA tool call change called a wrong API or something that unintentionally allowed any email verification attempt that was successful to be used, but the engineers did not add a sufficiently thorough e2e test case to test the tool call against unrelated email verification attempts being provided to the tool call. This is the part I think should be focused on the most. Tool calls for agents that have their output potentially influenced by an attacker should be treated like external APIs that anyone can reach, and they should be tested as such.
This is all obviously a guess, doesn't take into account the many signals they use to determine if an account recovery attempt is valid, and could be very inaccurate, but it's the closest to what I (someone who deals with Meta security a lot) think could have allowed this to happen.
Some Jr engineer got tired of handling stupid support requests and automated the job with an agent. That’s how.
Assigning Jr engineers for security support is ridiculous partly because young people don’t understand how critical security is sometimes. And partly because they don’t value privacy as much.
> But it should only be able to "hit a button" to send a 2FA email to the address attached to the account, all run with hand-written code.
Genuine question...why would that need to be hand-written?
It makes absolute sense as a general statement and is kinda crazy that this wasn't a built-in limitation, but I'm not quite sure why the code for that bit must be hand-written (provided the code functionally does what you describe).
I think he likely means "code that is hand-reviewed" and not directly controlled by the agent. He's probably meaning to differentiate it against the in-process agent writing the code. It doesn't matter too much if that fixed code was written by an LLM under guidance and review of the SWE, outside the agent.
Maybe not hand-written, but definitely static, and at least human-reviewed/tested to only allow sending to previously-validated email addresses.
This exploit is my new gold standard for trivially avoidable security failures. Someone has finally beaten Gitlab's password reset emails to attacker-provided addresses.
The harness is vibe-coded.
I'm among the first 6000 users of Instagram and my first name username was stolen a few years ago. Support for verified accounts acknowledged the issue, but couldn't do anything about it.
This turn was an AI exploit, in my case was an outsourcing support 'exploit', where someone paid for my username to be manually changed and given to another user. There will always be a way to get access to accounts if human accountable support doesn't exist, with criminal consequences for employees that violate it.
> with criminal consequences for employees that violate it
lol, no. The day someone is criminally charged with "stealing" a username is the day that humanity has lost
Support requests have always been the weakest link in the security chain for big corps. I've had accounts of mine turned over with 2FA disabled by humans before. I guess we shouldn't be surprised that the LLMs are doing the same thing.
The simple fact that 2FA can be removed by low level support staff drives me mad. It defeats the whole purpose of the process.
>> The simple fact that 2FA can be removed by low level support staff drives me mad. It defeats the whole purpose of the process.
The fact it can be removed by anyone is the problem. If you lose access to your 2FA (and recovery codes) then you should lose access to your account. Having it removable by anyone (other than a logged in account holder) defeats the entire point.
100%
Urgency.
Emotions.
It's all there, and high-stakes environments with no proper protocol are most vulnerable.
Source: used to work part-time in IT support at a hospital, by now 10+ years ago, so it was routinely requested to circumvent regulations and security protocols, even medical ones (cough Windows in ICU monitors and other medical "kiosk" PCs that should absolutely not run Windows)
I love those admin passwords which a tech will give you at some point because he doesn't want to do the work himself. If they even have passwords...
Unfortunately Siemens woke up.
recovery is always the weakest link in any authentication system
This is not wrong but what’s really missing is cost: Meta did this so they can avoid paying people to do it. Lots of companies follow that decay spiral: your bank could shut phishers down cold by requiring wire transfers to be authorized in person but they don’t want to pay staff or risk you being upset by a transaction taking an extra hour so they don’t.
Imagine an alternate universe where big tech companies worked with various trustworthy third-parties where something like this would generate a challenge you could take to your local notary, post office, library, police station, etc. where someone would check ID before approving it. How many phishing attacks would be prevented annually by a physical presence check?
> your bank could shut phishers down cold by requiring wire transfers to be authorized in person but they don’t want to pay staff or risk you being upset by a transaction taking an extra hour so they don’t.
Isn't this essentially what just recently happened to the Pope? Then there were people here doing the rest of your comment for him saying how egregious it was for them to ask for an in person authorization. It sounded like all he was trying to do was update his address, but changing your address from one in Chicago to one in a European country absolutely sounds like something a phisher would be trying to do.
Then you get trusted parties selling account access. Even if you remove them for a single false positive they will do it. A bit like a % packages "vanishing".
The least terrible seem digital id.
for a while facebook had the ability to recover your account by having them ask several of your friends if the recovery was legitimate but it was turned off. my guess is that not enough people added trusted contacts to bother running it.
https://www.theverge.com/2013/5/2/4292744/facebook-trusted-c...
The amount of hassle involved with regular physical checks is why it's not implemented, regardless of attack prevention.
The cost of hiring a person is part of it but not really the core reason. People were sold on the Internet with "you can do things online conveniently" and reintroducing the need to physically go somewhere negates that angle entirely.
> People were sold on the Internet with "you can do things online conveniently" and reintroducing the need to physically go somewhere negates that angle entirely
But how often does one need to do recovery procedures like this?
How much less convenient is it for everyone else to be at risk of their account being taken over?
It's a hard problem. How do you prove you own an account if you lost all proof of ownership? Especially so if an account was never tied to your real name, in which case you could at least rely on government ids.
It's a tough problem, because people forget passwords, change phones, lose access to 2FA devices, but still need to use their accounts.
I had to go through the account recovery on my Facebook account once and the proof they demanded was that I match a bunch of pictures of friends to their names. I think it took 3 tries over multiple days to actually get it unlocked because it turns out I such really remember a lot of the people I met 20 years ago and friended on Facebook.
I don’t recall why I had to go through this song and dance. Very plausibly the account was still associated with an old school address that I could no longer access. So yeah, account recovery is hard. How do you prove someone owns an account when they’ve lost the things they are supposed to use to prove ownership?
I manage customer identity and access management ("CIAM") for a financial services firm. Passkeys are primary, recovery can be performed by providing a government credential remotely (which costs us ~$2-3 per recovery). I do not think it is hard, based on what we have built and spent to enable these capabilities. NIST Special Publication NIST SP 800-63 Digital Identity Guidelines is a helpful resource on this topic.
https://pages.nist.gov/800-63-4/
I think Meta just does not care if they're enabling AI attack surface and vulnerabilities into these customer journeys. It's...certainly a choice, versus deterministic journeys with hard guardrails. They could make different choices.
> recovery can be performed by providing a government credential remotely
That only works because you presumably do KYC when you open accounts, so you have an identity to match to. Most internet accounts don't do real KYC, so a government credential doesn't really work for recovery --- they didn't know who you were, so proving who you are doesn't help anything.
That doesn't mean that letting anyone sweet talk support or an AI into taking over an account is acceptable, of course.
It's a fair point, and can be solved for as part of the "Verified" offerings Meta offers. This binds IRL identity to the digital identity at verification for future identity assurance step up (including if and when recovery is required). Failing that, TOTP, SMS, and even mailing an OTP to a mailing address remain low friction auth factors (with, of course, various levels of security).
My point is that while this is not easy, there are obvious very bad ways to implement this that should not be done (chatbot or other generative AI interface vulnerable to the usual suspects of AI inherent attack surface). Don't build the bad way, the right away is known and straightforward.
I’d wager your range of tech literacy/capabilities for your firm is much narrower than big tech.
Range != value, depending on use case. Doing more poorly does not make something better. Our customer identity capabilities are very close to login.gov (we don't have to support hundreds of agency customers and common access cards), and if its good enough for ~342M Americans, its good enough for our customer base.
Broadly speaking, work for the sake of work is not valuable work. Show me outcomes for resources and time invested, and compare accordingly. Value is, again broadly speaking (there is always nuance), what you deliver. If you bring me an AI solution for a high risk high value customer journey, data flow, or code path, that is an anti pattern. If you, as a colleague or a stakeholder, put forth that we must use AI in situations that require a high degree of determinism (due to potential high cost failure modes), you will need to prove this extraordinary claim with evidence.
Choose Boring Technology - https://news.ycombinator.com/item?id=9291215 - March 2015 (212 comments) ["Am I using this project as an excuse to learn some new technology, or am I trying to solve a problem?"]
I get paid to manage risk efficiently, including being measured on time and budget spent against the success criteria, ymmv; my comp and budget is not dependent on how much AI I shove into security systems. "What am I optimizing for?"
Amazon scraps AI leaderboard to stop workers chasing usage scores - https://news.ycombinator.com/item?id=48315583 - May 2026 (19 comments)
fair enough, but what's the actual point of 2FA if it's so easy to override?
the alternative is people losing their accounts and people aren't willing to allow that. i do think that apple does this a little better where they try everything to contact you in every way they know and it takes a week to get access. at a minimum to change your email it should require a week of waiting to see if the user can access the original mail to the hand off.
It depends. Some like AWS take it deadly seriously and it takes a long time to recover root access to an account.
So the AI agent had privileged access to remove 2FA, ignore the account email, and just hands accounts to whoever asked? Honestly that’s so highly negligent I wonder if the implementation team for that “feature” was intentionally trying to do as much subtle damage to meta as possible before their inventible layoff.
It’s a shame nobody tried to get it to drop the production table entirely! (mostly joking). Just claim to be a high level SRE solving some critical production bug, the only solution to which is dropping the database.
I get that account recovery for sites with hundreds of millions of users is a huge burden they're struggling to manage but I'm shocked they didn't restrict such loose verification to the >90% of lower value accounts that aren't worth stealing and keep the stricter verif on high-value accounts.
The next obvious thing would be to let accounts the algorithm judges to be low-value still opt-in to strict verif. The vast majority of low-value accts won't bother flipping it on if the option is buried two menus deep, but many of the few low follower/views accts who are targets for some other reason (political, stalker, etc) - know they are targets and can self-protect by opting in, further reducing account hijacks.
So, before we even get to whether this 'loose' verif is "bad", those two simple implementation changes would certainly have cut the bad outcomes of a (potentially) bad idea by >95%.
This is true for any service that Meta owns. I experienced something similar on my Meta (formerly Oculus) account. Meta support is very susceptible to social engineering and they have been for some time.
Always a bit illuminating to me how many exploits seem to so dumb I'd never even bother to attempt them. You're telling me I can just...ask for the password? And that works?
It's not called artificial intelligence for nothing.
This is an embarrassing failure for Instagram. But SIM cards have been hacked the same (by tricking support, claiming the phone was lost or stolen), except the agent was human.
The solution (which also solved SIM support agents being bribed or hacking known acquaintances) was to prevent the agents from resetting the SIM card without some steps the original owner would have to follow (and could follow even if they've lost their original phone), like a PIN they'd have to remember. I think the same solution should be applied to AI agents.
>Once it looks like the request is coming from the correct region, they tell the Meta support AI that the account is hacked and ask it to send the verification codes to an arbitrary email address they control.
Dear Instagram, wtf. Why not send the reset to the account in question? Arbitrary email, wow.
Perhaps the attacker says that they email was also hacked and "this is my new email now". It sounds like this was a result of AI support and not a real person "And if you're part of the A/B tested accounts on which the AI support option is active, tough luck, you can't even turn it off."
The implications of this are quite unsettling. Meta gave an agent privileged read AND write access to user accounts with no human in the loop?
Yep... And just think: this is what AI boosters want us to do.
Yes. AI is in charge now
> with no human in the loop
With no basic validation either apparently. Insane.
Disgraceful. Instragram's "security" has been trash for years.
How is this "embarrassing" instead of subject to legal liability?
We really need similar rules to other engineering disciplines. If your building falls with people inside, you killed them.
Nobody dies if instagram collapses. Might even cause more people to live.
You said it, instagram is not life-critical
This happened to my instagram yesterday night while I was asleep. I don't have a particularly high value username (it's probably worth somewhere in between $300-500), but still incredibly frustrating to deal with. True to the article, I had already enabled 2FA last night and it didn't matter.
Thankfully, IG gave me the option of restoring my username when I logged back into my account today.
The ironic thing is I know several legitimate humans who have lost access to their accounts and have been trying for months/years to get restored access, but being stuck in support hell.
Maybe they should have hacked themselves.
I'm sitting here wondering why the Chief Master Sergeant of the U.S. Space Force has an Instagram account to begin with. I understand it's the office itself, but still don't see the reason to expand the attack surface of government offices. X makes sense, Instagram, I'm not so sure as much
Outreach, I'd guess? You've got to do outreach where the people are. X and Instagram have pretty different audiences, but they're both large, so if you're on one you probably should be on both.
I see no difference between X and Instagram in this regard whatsoever.
Think NASA, for example; it's also a government agency, and they are doing great job posting photos in Instagram, do you think anything is wrong with it?
They're just one tiny step from the AI emailing itself all the account recovery links, and locking out the entire userbase.
It might even do that preemptively if it thinks they're going to shut it down.
Nothing says you are an advanced stupid company than using AI to implement the stupid. This is security I doubt even a college student would implement. Does Meta have a CSO? The correct answer is they don't, even though some body might occupy the title.
Of course it's always possible that they simply don't care who has your account, as long as they get money.
Is there any credible primary source for this exploit being real?
https://www.404media.co/hackers-simply-asked-meta-ai-to-give...
Related discussion: https://news.ycombinator.com/item?id=48350239
wow thats extremely embarassing for meta
Just another day for Meta in terms of embarrassing outcomes, and yet the company makes hundreds of billions of dollars per year because the only thing that matters anymore is shoving increasingly scammy and worthless ads in front of as many eyeballs as possible, even when the people with those eyeballs can less and less afford to buy anything non-essential.
I suppose you could chalk this up to an oversight. I don't see how Meta gained from this. They've been purposeful about collecting user data and lying about it, eg: 2025 Android Tracking Incident. Shouldn't just be an embarrassment, should be much worse than that.
Who specifically do you think is embarrassed there? They’ve got all the cards, they don’t care.
What is even the point of having 2FA if it can be so trivially bypassed? Isn't that the whole point that it's sort of a last line of defense? Oftentimes, you can't change simple account settings without having to re-auth and then punch in your code again. Why would something as critical as a suspicious password reset be able to jump ahead of that? Mind boggling. But, I guess that's what happens when you lay off 10% of your people at a time.
If the LLM has knowledge of something, by design it can't help but divulge it. When will companies learn granting any kind of sensitive information access to an LLM is a moot point
What part of this article implied the LLM divulged sensitive information to a user? All it did was change your associated email if you impersonated the user
I think the related news of Meta rolling out subscription models for their free products, is a step in the right direction.
Otherwise the only way to provide these services is to massively underfund support, if you charge 0$ per account and serve 1 Billion users, then you cannot afford to spend 1 minute of human support time on an account.
Yes, they could use the money from ads, but let's be frank, the customers in that case are the sponsors, if the customer is the actual user, then it's way easier to provide direct support to them without facing an foundational incentive misalignment.
The only thing worse than a naive customer support rep is an even more naive customer support ai.
Jeez, straight up amateur shit. Genuinely hard to believe.
"Social engineering is all you need"
More like "Prompt engineering" ?
Can we really name this "Prompt engineering"? The prompt is so simple this is hardly any work even less than this comment
My account, with a 3-letter username worth $$$, got hacked yesterday morning probably by this flow, but I did manage to defend it. I think by far the biggest problem with Instagram/FB/Meta auth flow is that 2FA does nothing. You don't need the 2nd factor to disable it, so attackers can just turn it off. Really stupid!
Also, I discovered that many of IG's auth endpoints are just broken. For example you can't change password on web because of CORS, which isn't a transient outage but just a flat out bug.
Edited to add: This is just the cherry on top of years of stupid auth flow at IG. I have received tens of thousands of reset links or codes from IG over the years. There used to be a way to put your account on recovery cooldown for a few weeks but they got rid of even that.
I'm horrified with how poor Meta's use of AI is recently. Here's a list of the issues both me and my wife have been plagued with over the past few weeks. It's really quite an achievement to be this terrible. 1. My personal Facebook received 3 violations restricting my ability to manage ANY Page until April 2027 (lol). The trigger... I deleted 3 unused Pages. These Pages I had created years ago in preparation for projects that never came to fruition, and had never posted any content. THe pages were 'scheduled for deletion', and when that day came (around a month later?), boom, I'm hit with a 1 month restriction which later converted itself into a 1 year restriction after I waited out the month. No Appeal button. I'm expected to wait for a year to manage my new page? All over something that is NOT a violation, just for deleting old pages. Get out of here. Smart system.
2. I pay for Meta Verified on Instagram and for the past 2 weeks "Enhanced support" leads me to a broken interface. "Page isn't available right now". So, what am I paying for exactly?
3. It seems you can use Meta's AI Assistant to sometimes get through to a human. I've done this twice now, and both times my case has been escalated to a different team (apparently) yet I never get an email, I never get an update in the chat (the chat ENDS immediately after the phone call with support), and the issue is never resolved. It's been 2 weeks. The case says "Completed", with no response. Worthless as always.
4. My wife creates content on Instagram and has had her account suspended multiple times now for "Account Integrity". I assume the system thinks she's not the person in the content, despite providing her valid email, phone number, video selfie, and 2 types of ID (passport & driver's license) multiple times. What's hilarious is the passport was accepted on of her accounts (they wiped out everything on her Account Center), but another account was rejected. Great AI, same passport, exact same lighting... different outcome.
So as it stands, we're both fucked on both facebook and instagram thanks to awful AI moderation, and fucked further thanks to awful AI support. No resolution in sight. The incompetence is next level. I really don't see this getting resolved. This already happened to my wife earlier in February, she managed to get one account back, and a month later she's hit with the same identity issues.
Using AI for both the moderation and the support makes me sick. The same poor AI that incorrectly flagged me and my wife's accounts for a load of incorrect bullshit is the same system that's meant to help resolve it? Of course it's going to side with its own poor decision. YouTube seems to do the same thing and auto-reject appeals in seconds. Really smart /s
I believe we need enforcement that social platforms should NOT be using AI to perform destructive actions without human intervention. Noone should ever lose their accounts because of AI mistakes. AI should be used to surface potential issues which get passed to a HUMAN to double check before applying the action. AI simply isn't good enough to have full control.
Fucking pissed off and even angier now I've had to write all this up and remind myself just how ridiculous the situation is. Sorry for the rant, but losing your accounts you put work into is very crushing and demotivating. Being accused of these violations fills us both with so much resent for the companies running this shit.
Sam Cofounder Postmates
On the off-chance there's anyone at Meta seeing this (@Wirah on twitter)
Had to make this new username as my original (samstr) comment doesn't show up. No idea why. Probably shit AI
But I was told that when Zuckerberg bought IG, it wasn't to murder competition in its crib. Instagram "only had 12 employees" so it must be ok
It sounds really insane. Too bad there is 0 proof or anything in the article, so I am very skeptical. Without proof etc this is just a very nice doom story.
The proof is that you Google this right now and find multiple corroborations across the web from today.