Blog/Incident analysis

The Snowflake customer breaches: 165 companies, no platform vulnerability

Mandiant traced data theft at around 165 Snowflake customers to stolen credentials, no MFA and years-old infostealer logs. The platform was never breached. Here is what that means.

CyberWatch AI21 November 2026 · 7 min read
A server cabinet with illuminated status lights and patch cables in a dim equipment room

In the summer of 2024 a series of very large data breaches were reported as "the Snowflake breach". Ticketmaster. AT&T. Santander. Advance Auto Parts. Neiman Marcus. The common factor was real, and the name was wrong in a way that matters.

Snowflake was not breached. Its platform was not exploited, no vulnerability was involved, and the company said so jointly with CrowdStrike and Mandiant: no evidence that the activity was caused by a vulnerability, misconfiguration or breach of the Snowflake platform. What happened instead is that an attacker logged in, hundreds of times, with correct usernames and passwords belonging to other people, and nothing stopped it.

Attack, enabled by configuration

This was a criminal attack, not an error or an outage. Mandiant tracks the actor as UNC5537 and published its analysis in June 2024. Snowflake says it became aware of potentially unauthorized access to certain customer accounts on 23 May 2024, with increased threat activity observed from mid-April.

But the reason the attack succeeded at this scale sits in customer configuration, and the findings are unusually specific about it. Mandiant reported three conditions across the affected customer instances:

  • No multi-factor authentication required. The affected instances did not enforce it.
  • Credentials not rotated. In many cases the username and password had been unchanged for as long as four years.
  • No network restrictions. The accounts were reachable from anywhere, with no allow-list limiting which networks could connect.

Any one of those would have made the attack harder. All three together make a stolen password sufficient, permanently, from anywhere.

Where the passwords came from

The most uncomfortable detail is the age of the credentials. Mandiant found that the majority of those used by UNC5537 were available from historical infostealer infections dating back as far as 2020, naming families including Vidar, Redline, Raccoon, Lumma and Metastealer. Roughly 80% of the accounts the actor used had prior credential exposure in that ecosystem.

Infostealer malware does one job: harvest everything a browser has saved and send it to a collection server, where it is bundled and sold. A contractor's laptop infected in 2021 can produce a working corporate data warehouse login in 2024, because nobody rotated the password and nobody turned on a second factor. Snowflake's own statement noted a related thread: a threat actor obtained personal credentials to demo accounts belonging to a former employee, accounts which Snowflake said did not contain sensitive data.

That is the mechanism in full. No exploit, no malware on the target, no lateral movement. A login.

What was taken

Mandiant and Snowflake said they had notified approximately 165 potentially exposed organizations. That figure is a notification count published by the investigators, and it should not be read as 165 confirmed losses.

Two victims dominate the public record. Ticketmaster's parent Live Nation disclosed unauthorized activity in a third-party cloud database environment containing company data, identified in May 2024. AT&T disclosed in July 2024 that call and text records covering nearly all of its cellular customers from 1 May to 31 October 2022, plus a small number from 2 January 2023, had been illegally downloaded from its workspace on a third-party cloud platform. The figure reported from AT&T's disclosure was approximately 109 million customer accounts. AT&T said the files did not contain the content of calls or texts, Social Security numbers or dates of birth, and noted that time stamps were not included, while also acknowledging that publicly available tools can often link a phone number to a name.

That caveat is the heart of why metadata breaches matter. A complete record of who called whom, for how long, across a hundred million people, is not neutral because it excludes the conversation. It is a social graph.

Confirmed versus unknown

Confirmed: the absence of a platform vulnerability, the credential origin in infostealer logs, the missing MFA and network restrictions, the 165 notification figure, the dates, and the identity of the actor cluster. Confirmed in a legal forum rather than only a research one: the US Department of Justice unsealed charges in November 2024 against Connor Moucka and John Binns, alleging intrusions at at least 10 organizations and that the pair extorted at least 36 bitcoin, worth roughly 2.5 million dollars at the time of payment, from at least three victims. Moucka later pleaded guilty in Seattle federal court.

Not established publicly: the full victim list, the total volume of data taken across all 165 organizations, how many paid, and which specific infostealer infection produced which credential. The gap between "at least 10 organizations" in the indictment and "approximately 165" in the investigators' notification count is a difference between what prosecutors chose to charge and what responders observed, not a contradiction, and it is worth not conflating the two.

Why it mattered

It redrew the line in the shared responsibility model for data platforms. Everyone agrees the cloud provider secures the infrastructure and the customer secures their configuration and identities. In practice, when a data warehouse is breached through an un-MFA'd login, the headline says the provider's name. Snowflake was correct on the facts and lost the narrative anyway, which is a lesson for every SaaS vendor: if a default makes customer compromise easy, the resulting breach will be reported as yours. Our cloud security guide works through where those boundaries actually fall.

It made optional MFA look indefensible. At the time, Snowflake applied MFA per user, with users expected to enrol themselves, which meant an administrator had no simple way to guarantee coverage across a tenant. An authentication control that depends on individual voluntary adoption is not a control. The industry direction since has been towards administrator-enforceable MFA and, for service accounts, key pair or OAuth authentication instead of passwords.

It demonstrated the long tail of infostealer logs. Credential theft from an employee's browser is usually filed as a nuisance. This event showed those logs functioning as a durable inventory of enterprise access, queryable years later. If you treat infostealer infections as a cleanup task rather than a credential-rotation trigger, you are leaving usable keys in circulation.

What to take from it

  1. Enforce MFA on every data platform at the tenant level. Not recommended, not available, enforced. Then verify coverage by listing users without it, which Snowflake and most comparable platforms let you query.
  2. Use network allow-lists on anything holding bulk data. A network policy restricting a warehouse to your corporate ranges and known integration IPs defeats a stolen password outright, and it is one of the specific controls Snowflake's own guidance highlighted for highly privileged users and service accounts.
  3. Treat service account credentials as the real problem. They are the ones without MFA, without rotation, shared between teams and embedded in scripts. Move them to key pairs, scope them narrowly, and rotate on a schedule somebody owns. This is least privilege applied where it is least convenient and most valuable.
  4. Rotate credentials after any infostealer detection, including on personal and contractor devices. The infection is cheap to clean. The credential exposure is the lasting damage, and it does not expire on its own.
  5. Hunt in your own logs. Snowflake and Mandiant both published detection guidance and queries: look for logins from unfamiliar client applications, large result exports and unusual user agents. The capability to answer "did this happen to us?" within a day is worth building before the next advisory lands. Our threat intelligence guide covers turning an advisory into a hunt.

There is also a procurement lesson, and it is not about Snowflake. When you put a large volume of regulated data into a managed platform, the questions that matter are not about the vendor's certifications. They are about which authentication options the vendor allows you to make mandatory, whether you can restrict access by network, and whether you can see every login. A vendor that cannot let an administrator enforce MFA has handed you a risk you cannot manage, however good their compliance documentation looks. That is a third-party risk question, asked before signing rather than after a notification.

Told your data was in one of these breaches? Extortion and fake-notification emails followed these incidents closely. Paste the message or link into CyberWatch AI Scan for a free check before responding.

One final observation about naming. Calling this "the Snowflake breach" was convenient and it misdirected a lot of remediation effort, because it implied the fix belonged to a vendor. The fix belonged to 165 security teams, each of whom could have prevented their own instance of it with a setting. The uncomfortable version of the headline, which is also the accurate one, is that an attacker walked through the front door of a hundred and sixty-five companies using passwords those companies had allowed to remain valid for four years.

Frequently asked questions

Was Snowflake hacked?

No. Snowflake, with CrowdStrike and Mandiant, stated it found no evidence that the activity was caused by a vulnerability, misconfiguration or breach of the Snowflake platform. The attacker logged in to customer accounts using valid stolen credentials.

Then what actually happened?

A threat actor Mandiant tracks as UNC5537 used usernames and passwords harvested by infostealer malware, in some cases years earlier, to log in to Snowflake customer instances that did not require multi-factor authentication, then exported and extorted the data.

How many organizations were affected?

Mandiant and Snowflake said they had notified approximately 165 potentially exposed organizations. That is a notification count, not a confirmed-loss count.

Which breaches came out of this?

Ticketmaster and AT&T were the most prominent. AT&T disclosed that call and text records of nearly all its cellular customers from May to October 2022 had been downloaded from its workspace on a third-party cloud platform, reported as approximately 109 million customer accounts.

Why did multi-factor authentication not stop it?

Because it was not switched on. Mandiant reported that the affected customer instances did not require MFA. At the time, Snowflake enabled MFA per user and relied on users enrolling themselves, so an administrator could not easily guarantee coverage.

How old were the stolen credentials?

Mandiant found that the majority came from historical infostealer infections dating as far back as 2020, and that in many cases the credentials had not been rotated for as long as four years. Roughly 80% of the accounts used had prior credential exposure.

Was anyone charged?

Yes. The US Department of Justice unsealed charges in November 2024 against Connor Moucka and John Binns, alleging intrusions at at least 10 organizations and extortion of at least 36 bitcoin from at least three victims. Moucka later pleaded guilty.

What changed at Snowflake afterwards?

Snowflake moved towards letting administrators require MFA rather than leaving it to individual users, and published hunting guidance and queries for customers to check their own accounts. The broader industry shift was towards treating password-only access to a cloud data warehouse as a configuration defect.

Sources

  1. UNC5537 Targets Snowflake Customer Instances for Data Theft and Extortion, Google Cloud Threat Intelligence
  2. Snowflake Recommends Customers Take Steps to Prevent Unauthorized Access, US Cybersecurity and Infrastructure Security Agency
  3. Detecting and Preventing Unauthorized User Access, Snowflake
  4. Unauthorized User Access to Snowflake Customer Accounts (AL24-008), Canadian Centre for Cyber Security
  5. Threat hunting guide for Snowflake environments, Google Cloud / Mandiant
  6. Mandiant: Exposed credentials led to Snowflake attacks, TechTarget
  7. AT&T says hackers stole 2022 call and text data from 'nearly all' cell customers, NPR
  8. Canadian pleads guilty to Snowflake cloud data theft attacks, BleepingComputer
  9. Canadian Man Pleads Guilty in Snowflake Extortions, KrebsOnSecurity
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.