Skip to main content
All incident reports
Incident report

Cloudflare fixes Containers flaw that let one tenant read another's leftover disk data

A misconfigured shared storage pool in Cloudflare Containers and Sandboxes returned deleted containers' disk blocks to other customers without wiping them first. Cloudflare fixed it in September and says it found no sign that anyone but the reporting researchers used the method.

· First published

The short version

  • The flaw. Cloudflare Containers built each container's root disk on thin-provisioned Linux storage, handing out space in 64-kilobyte blocks. When a container was deleted, its blocks went back to a pool shared across customer accounts, and that pool was set to skip wiping a block before reuse.
  • What could be read. By writing 4 KB into a reused block, a customer on a Workers Paid plan could read back the remaining 60 KB at raw disk level. Recovered material included directory structures, database pages, complete SQLite databases, browser profiles, .env files and credential files, according to the researchers.
  • The fix. Cloudflare turned wiping back on, then retired every running container disk and cleared cached image layers, finishing on 19 September. It disclosed the flaw five days later and says the changes were applied to its infrastructure automatically, so customers need do nothing.
  • No evidence of abuse. Cloudflare built detection signatures from the researchers' proof of concept and its own copy of the attack, ran them against retained disk-activity records, and says it saw only the researchers' and its engineers' authorised testing. It did not state how far back those records go.

Timeline

DateEvent
4 SeptemberOren Yomtov of the security firm Accomplish reports the flaw through Cloudflare's bug bounty programme.
14 SeptemberResearchers confirm their proof of concept no longer works after Cloudflare re-enables block wiping for newly handed-out blocks.
19 SeptemberCloudflare finishes retiring running container disks and clearing cached image layers.
24 SeptemberCloudflare discloses the flaw, five days after completing cleanup.

How the cross-tenant read worked

Cloudflare Containers runs customers' programs in containers on servers shared by many accounts, and Cloudflare rather than the customer chooses the host. Each container's root disk was built on thin-provisioned Linux storage, which allocates space in 64-kilobyte blocks. When a container was deleted, its physical blocks returned to a pool that served workloads belonging to multiple customer accounts.

That pool was configured to skip zeroing a block before handing it to the next container, although wiping is normally the default. A new container that wrote only 4 KB into a reused block therefore overwrote just that portion, leaving the other 60 KB holding the previous tenant's bytes. Reading the whole block back at raw disk level exposed them. Cloudflare Sandboxes, sold as a place to run untrusted and AI-generated code, ran on the same service.

What the testing found

The researchers reported finding leftover material on 18 of 24 container placements and on 20 of 22 underlying machines across four continents, all on servers Cloudflare selected. The blocks held directory structures, database pages and structurally complete SQLite databases; the researchers' own write-up adds Chromium browser profiles, .env files and credential files, and describes the material as other customers' files.

Neither source shows that a tenant could influence whose data it received, read an actively attached disk, alter another customer's live data or take a workload offline. The researchers say their analysis scripts printed only counts and format checks rather than file contents, and that the material they sent Cloudflare carried no third-party names, identifiers or credentials; Cloudflare says that data was kept private and securely deleted.

Cleanup, disclosure and an unresolved scope question

Cloudflare's first change, re-enabling wiping for newly handed-out blocks, did not clean blocks already mapped into running container disks or into each server's cache of prepared image layers, which a new container could inherit. The company therefore drained and restarted servers during quiet hours to retire every running container disk and clear those caches. It then searched its retained disk-activity records and reported finding no use of the method beyond authorised testing.

The two accounts differ on scope. Cloudflare's post named Containers and Sandboxes as affected and did not mention Browser Run, while the researchers say the same disk setup also affected Browser Run. Cloudflare also did not say when the unsafe setting was first put in place, so the length of the exposure is not established by either source.

What is not known yet

  • When the non-wiping setting was introduced and how long customer blocks could be read before the fix.
  • Whether any customer outside the researchers' testing actually recovered another tenant's data.
  • Whether Cloudflare Browser Run was affected, as the researchers say and Cloudflare's own post did not address.
  • How many customers or containers ran on hosts using the misconfigured pool.
  • Whether any credential material left in a reused block belonged to a still-valid account.

What to do

If you use Cloudflare Containers or Sandboxes

  1. Take no remediation action for the flaw itself: Cloudflare says the fixes were applied to its infrastructure automatically.
  2. Read Cloudflare's advisory if you need the affected product list for an internal record or audit.
  3. Treat the service as still suitable for untrusted code, but keep in mind that the issue exposed files left behind by other tenants rather than your own live workloads.

If you run multi-tenant platforms on shared or thin-provisioned storage

  1. Confirm that blocks are zeroed before they are reassigned to a different tenant's volume.
  2. Check the pool-level configuration rather than assuming the default wipe behaviour is still in force.
  3. After correcting the setting, also retire existing volumes and clear cached image layers, because the configuration change alone leaves already-mapped blocks readable.

For everyone else

  1. When a storage misconfiguration is found in shared infrastructure, verify that the fix covers data already allocated and cached, not just future allocations.

Update log

  • First published.

Sources

New incident reports by email

One email a week with the reports published since the last one.

Want us to look at your domain?

Give us one apex domain and we'll send back a written snapshot of what is already visible from outside. No call required.