Skip to content

Product Not yet confirmed elsewhere1 publisher 2 min read Published

AWS quarantines leaked keys. Truffle Security's sweep says 9,342 of them still work.

A sweep of public code, datasets and CI logs found 768 AWS keys with full account control. The containment policy AWS applies to detected leaks still permits deleting the audit trail.

The Product Desk

Send a correction

What happened

  • Truffle Security gathered 431,875 AWS secrets from repositories, git history, datasets, Docker images and CI logs, and reduced them to 64,024 unique keys.
  • 768 of the keys granted full control of a corporate AWS account, 526 of them root credentials.
  • The median exposed key with a known creation date was about five years old.

Why it matters

  • What doesn't add up contradiction Amazon presents quarantine as containment; Quinn's reading of the same published policy describes a credential that can still delete the account's audit trail, so the two accounts of the control...
  • What it limits constraint Because the policy is written not to disturb existing resources, there is a ceiling on how much it can ever restrict, which pushes the real burden onto rotation.
  • What it costs cost Compliance-mode retention set by a quarantined key produces a storage bill the customer cannot cancel by asking support, only by destroying the account.
  • Who's exposed exposure A five-year-old root key sitting in a public dataset is exactly the third-party technology risk DORA obliges European financial entities to identify and document, and the obligation has been live...

The policy Amazon applies when it spots a leaked key is meant to limit fraud-related damage without, in its own phrasing, impacting existing resources [8]. That second clause does most of the work. Writing in The Register, the cloud economist Corey Quinn reads the published policy and lists what a quarantined credential can still do: assume other roles in the account, run commands on running instances, stop CloudTrail logging, and delete the audit trail outright [14]. Quinn also notes that writing to a bucket is permitted, as is setting object lock and retention, so an attacker can fill storage and apply compliance-mode retention that nobody can shorten, AWS support included, without deleting the whole account [15].

That is one practitioner's reading of a document rather than a demonstrated attack [16], and it deserves that label. It is also checkable line by line against Amazon's own documentation, which makes it harder to dismiss than a proof of concept nobody else can reproduce.

Now the arithmetic of the sweep itself. Truffle Security could only fully test 10,616 of its 64,024 unique keys, or 16.6 percent of the corpus [10]. Inside that slice, 9,342 credentials still authenticated [9], and roughly 68 percent of the full-control finds were root keys rather than lesser IAM users [11]. The 768 figure is therefore a floor drawn from a sixth of the sample, not a census.

The rotation numbers are what make the quarantine policy load-bearing in the first place. Only 13.7 percent of users with an exposed key ever issued a newer one [6], which leaves 86.3 percent of that population with no successor credential at all [12]. A key with no replacement is a key that was never meant to be revoked, because revoking it breaks something.

The source mix is the part that should reallocate somebody's scanning budget. Hugging Face alone produced 8,482 unique key exposures, 13.2 percent of every distinct key in the study and the largest single origin [5][13]. TNW's framing is that model repositories have inherited the habits of software ones; the corpus also included Docker images and CI logs [4], which are not where credential hygiene programmes usually look.

All of which points at the same dependency. Quarantine only fires on keys Amazon detects as leaked [8]. The control that works without detection is rotation, and by Truffle Security's own count it is not happening.

What to watch

  • Whether AWS narrows the quarantine allow list, particularly the actions touching CloudTrail and S3 object lock, or restates that the current scope is intentional.
  • Whether Hugging Face publishes detection and revocation figures for cloud credentials committed to model repositories.
  • Whether any DORA supervisor cites long-lived static cloud keys in a third-party risk finding against a regulated financial entity.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence44
Adoption58
Hype gap+16
Incentives57
Confidence51
Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

  1. [1]

    Truffle Security found 768 leaked AWS keys granting full control of corporate accounts.

  2. [2]

    Among the exposed credentials were 526 root keys, the most privileged credential an AWS customer holds.

  3. [3]

    As of 10 August, 88% of the verified keys still authenticated, out of 10,616 tested.

Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.

Add to ChatGPTAdd to Claude

Sources

1 independent publisher whose own reporting we read for this story.

  1. thenextweb.com

    1 article · August 22, 2026

    Researchers found 768 leaked AWS keys that still work, and the containment policy leaves plenty possible

Share your take

Let Clarity write the post for you.

Signed-in readers get a short post drafted on this story in the register they choose — narrative, analytical, or a direct position — editable to the last word before it goes anywhere. The share buttons at the top of this story work without an account.

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Loading related stories