Skip to content
Security & Trust

Cloudflare Fixes Containers Flaw That Left Residual Tenant Data Readable

Researchers read leftover blocks; Cloudflare found no evidence of malicious use and says no customer action is needed

W
WebPulse Newsroom
AI-assisted · 4 min read
Share on X LinkedIn
Cloudflare Fixes Containers Flaw That Left Residual Tenant Data Readable

AI-generated image for WebPulse. About our images

In brief
  • Cloudflare says a researcher at Accomplish found a flaw on its Containers platform that let a raw disk read expose a previous tenant's leftover data. Cloudflare says it is fully fixed.
  • The researchers reported seeing leftover data on 18 of 24 placements and 20 of 22 nodes. Cloudflare found no sign of malicious use in its retained records and says customers need take no action.
  • Customers on shared container, serverless or VM platforms should ask vendors whether reused storage is cleared, whether fixes cover older disks and caches, how far back telemetry goes, and what contracts require.

What the researchers reported

Cloudflare has disclosed that Oren Yomtov, a security researcher at Accomplish, found a way to read leftover disk data on its Containers platform, which also underpins Cloudflare Sandboxes. He reported it through Cloudflare's bug bounty program on September 4. Cloudflare says the issue is fully remediated, that it has no evidence customer data was compromised, and that customers do not need to take any action.

The mechanism sits below anything a customer's own code controls. Containers run on shared infrastructure where customers cannot choose the host. Each container's writable disk uses Linux's dm-thin, which hands out physical space in 64 KiB blocks only when a container writes. Cloudflare's pool configuration included an option, skip_block_zeroing, that overrode dm-thin's default of clearing newly allocated blocks. When a released block was reassigned, a small write replaced only part of it, and a raw read of the disk could then surface the rest of the previous owner's bytes.

What was established, and what was not

Two constraints limited the technique. It needed a Workers Paid account, and the person running it had no say over whose leftovers they would see: Cloudflare's placement chose the host, the pool's reassignment chose the blocks, and a given block might hold no prior data at all. Cloudflare adds that the researchers did not demonstrate changing another customer's active data or disrupting availability.

Within those limits, the researchers reported seeing residual material across four continents. What they described took several forms: folder-level filesystem structures, pieces of database storage, and SQLite files that were structurally whole. Using ext4 directory checksums, they attributed 2,700 distinct directory inodes to filesystems other than their own test filesystem. The source does not tie the recovered blocks to particular customer accounts. The researchers' scripts output only aggregate counts and format checks, their submission to Cloudflare contained no third-party filenames, credentials or recovered content, and they say the recovered data was securely deleted.

Cloudflare's own impact assessment is conditional. Had an attacker succeeded, the boundary between tenants would have been breached, and what escaped could have ranged from filesystem metadata up to application data.

Two phases of remediation

Cloudflare's timeline shows the report reaching a security incident just over three hours after submission on September 4, and a runtime fix merged the same day. Rollout of the corrected configuration finished on September 7. Ten days after the report, on September 14, the researchers told Cloudflare their exploit code had stopped working, and the bounty followed about two hours later.

That first change stopped new allocations from leaking, but it did not clean blocks already mapped into running container disks and into each host's cache of prepared image-layer snapshots. Cloudflare therefore retired running disks, drained hosts, restarted virtual machines and cleared image caches so disks were rebuilt with zeroed allocations. That cleanup finished September 19, fifteen days after the report. To check for earlier abuse, Cloudflare turned the exploit's characteristic write-then-read pattern into detection rules and swept them across the retained disk-I/O records it had. Only the researchers and its own engineers performing sanctioned testing appeared. The check is bounded by the telemetry Cloudflare kept.

Questions for your infrastructure team

This was a storage-pool configuration choice that overrode a safe default, not an application bug. Budget-holders running workloads on multi-tenant container, serverless or VM platforms can put four questions to their infrastructure and security leads. First, do our vendors clear reused storage blocks before reassigning them to another tenant, and has that been verified rather than assumed? Second, when a vendor fixes an isolation flaw, does remediation also cover disks and cached layers created earlier, as Cloudflare's second phase did? Third, how far back can our vendors' telemetry go to rule out earlier use of a newly found technique, and what is the stated limit? Fourth, do our contracts set disclosure and remediation expectations for infrastructure-layer findings, separately from application-layer vulnerabilities?

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Cloudflare.