Skip to main content
Back
IntrusionAWS
9 min read

Amazon Bedrock abuse with enterprise AWS keys taken in a vendor breach

Victim
Unnamed organizations whose AWS keys for Amazon Bedrock were held by a breached vendor
When
Not stated; before Wiz's report of October 30, 2026Disclosed Oct 30, 2026
Entry
AWS keys for Amazon Bedrock held by an unnamed vendor, taken when the vendor was breached; the key type and how they were taken are not stated.

Wiz Research reported on October 30, 2026 that two enterprise AWS keys for Amazon Bedrock, held by an unnamed vendor, were first used within 30 minutes of the reported credential access in the vendor's breach. Days later, scripts probed the models and listed the quotas, and the keys ran about 800k Bedrock calls in about thirty hours, about 90% successful (Case A), and about 1.7 million in about thirteen hours, over 1 million successful (Case B), until each key was remediated. Case B's own key policy refused thousands of calls through global inference profiles, and regional quotas throttled nearly all of its calls that did not succeed, without ending the run; Wiz estimates over $100k in usage at list price.

What happened

External Parties
Resale operators
Breached vendor
Internet
AWS APIkey type not stated
Amazon Bedrock
Enterprise Bedrock Accounts
Amazon Bedrock Runtime
Bedrock and EC2 Quotas
Cross-Region Inference Profiles
Data Retention Setting
Vendor Integration Keys

Replaying the Bedrock vendor-key intrusion / start not stated; reported October 30, 2026

  1. 1

    Keys taken at the vendor

    Date not stated

    An unnamed vendor that holds the owners' AWS keys for Amazon Bedrock is breached, and the keys are accessed; how they were stored and taken is not stated.

  2. 2

    Keys tested within minutes

    Within 30 minutes of the reported credential access; date not stated

    Within 30 minutes, one model call from a residential address tests each key in a sweep across multiple organizations, and a single client then checks each key 14 to 120 times a day.

  3. 3

    Models and quotas mapped

    Four days (Case A) and seven to eight days (Case B) after the first test

    With the Case A key, a proxy pool probes the model catalog and scripts map the Bedrock and EC2 quotas across regions; with the Case B key, a checker finds the working US model ids.

  4. 4

    Bulk model use

    About thirty hours from four days after the first test (Case A); about thirteen hours from seven to eight days after it (Case B)

    The Case A key runs about 800k calls through global inference profiles; the Case B key, held to US profiles by its own policy, runs about 1.7 million; each key is then remediated.

  5. 5

    Account setting attempted

    Not stated

    An address tied to a model resale site, with access to both keys, tries to change the account's Bedrock data retention with the Case B key; whether it succeeds is not stated.

Impact

Two stolen keys ran Amazon Bedrock models at the owners' expense until each key was remediated: about 800k calls in about thirty hours, about 90% successful (Case A), and about 1.7 million in about thirteen hours, over 1 million successful (Case B), as Wiz published on October 30, 2026. Wiz estimates the usage at over $100k at list price; the actual billing impact is not stated.

Outcome

Where AWS controls fit for a vendor-held key

No AWS setting reaches how a vendor stores the keys you give it, or a breach of the vendor; that part is the vendor's. What you control is what such a key can do and from where. A service control policy that refuses Amazon Bedrock calls by the key's identity from any address but the vendor's would have stopped the key test and both bulk runs, if the vendor calls from fixed addresses and the account is a member account of an organization: Wiz places those calls at residential and hosting addresses and a shared VPN exit. A service control policy that keeps Bedrock Put, Update and Delete actions to approved administrators would have stopped the data retention change, if the identity behind the key is not an approved administrator, in a member account; whether the attempted change succeeded is not stated. The other rows limit a stolen key and do not stop it. A service control policy that refuses new long-term access keys and long-term Amazon Bedrock API keys, except for approved administrators, reaches no key that already exists; you would also have to delete the old keys, move the integration to a role and narrow the role's trust to the vendor. Service control policies that refuse Bedrock calls for unapproved models and through global inference profiles narrow what a stolen key can run. They are not a cap on loss. Case B's own key policy already refused the global calls, and the key still made over 1 million successful calls of about 1.7 million in about thirteen hours; the regional quotas slowed that run and did not end it. Approved models still answer a stolen key at the volume the quotas allow: the Case A key made about 800k calls in about thirty hours, about 90% successful. None of these binds the management account, and the sources do not say which key type was taken, when the keys were created, what the vendor's addresses are or whether the owners' accounts were member accounts of an organization. Wiz estimates over $100k in usage at list price, as published on October 30, 2026; the actual billing impact is not stated.

Could this have been prevented?

In hindsight, not as a whole. 4 controls would have refused a step outright, and 5 controls would have limited one. 1 step has no cloud-provider control. The counts below are control assessments, not attack stages: several can address the same step. The same replay with the controls in place before the incident shows where each phase would have been refused or bounded.

4
Prevented
5
Contained
1
Not covered

How it could have been prevented

The same replay with the controls in place. Green marks where a step is refused, blue where it is bounded, and red where no control reaches it. Open a row to read the control, why it acts, and its caveat.

External Parties
Resale operators
Breached vendor
Internet
AWS APIkey type not stated
Amazon Bedrock
Enterprise Bedrock Accounts
Amazon Bedrock Runtime
Bedrock and EC2 Quotas
Cross-Region Inference Profiles
Data Retention Setting
Vendor Integration Keys

Replaying the Bedrock vendor-key intrusion / start not stated; reported October 30, 2026

1 · Keys taken at the vendor

1 Contained1 Not covered3 on other clouds

2 · Keys tested within minutes

1 Prevented3 on other clouds

3 · Models and quotas mapped

1 Contained3 on other clouds

4 · Bulk model use

2 Prevented3 Contained6 on other clouds

5 · Account setting attempted

1 Prevented

Key facts

  • Wiz Research published the two cases on October 30, 2026, in relative time only, and its update notes through November 10, 2026 revise none of the counts; it names neither the vendor nor the key owners, and does not say whether the keys were long-term access keys or Amazon Bedrock API keys
  • Both keys were first used within 30 minutes of the reported credential access in the vendor breach, by one Amazon Bedrock model call from a residential address in a sweep of about one minute across multiple organizations; its first model failed and a fallback model succeeded, and a single client then checked each key 14 to 120 times a day
  • With the Case A key, a proxy pool probed the model catalog one call per model, and every model answered but one; scripts listed the Amazon Bedrock and Amazon EC2 quotas with both keys. Wiz reports the listing only, no use of Amazon EC2 and no access to another AWS service, and does not say that nothing else was reached
  • Case A's key, whose policy allowed global cross-region inference profiles, made about 800k Bedrock calls over about thirty hours, about 90% successful and about 6% throttled; Opus models were about three quarters of the successful calls
  • Case B's key was allowed only US cross-region inference profiles, and its own policy refused an apparent reseller pool's thousands of global-profile calls over four hours; one hosting address then made about 1.7 million calls in about thirteen hours, over 1 million successful, and regional quotas throttled nearly all of the rest without stopping the run
  • In late October an address tied to a model resale site attempted an account data retention change with the Case B key, and a second resale operation listed inference profiles and checked permissions on both keys; whether the change succeeded is not stated, and Wiz cannot definitively tie the resale operators to the initial compromise or the bulk abuse

Sources

Every fact in this breakdown comes from the public reporting below. The replay is a model of that reporting, not a reconstruction of the environment it describes.

  1. 1.
    "Black October" for AI Developer PlatformsWiz Research · Oct 30, 2026 · ResearchKey validation, daily verification traffic, model catalog probing, quota enumeration, bulk model use through inference profiles, the inference profile restrictions of both keys, call volumes, success rates and throttling, the data retention change attempt, the resale operators' key and permission checks, cost estimate
  2. 2.
    Northflank status page, security incidentNorthflank · Nov 6, 2026 · Status pageNorthflank's own security incident: a first known secret decryption between 02:39 and 03:08 UTC, a Wiz alert at 11:43 UTC and user reports of misused AI keys at 23:38 UTC on October 30, 2026, as its update of November 6, 2026 states; it does not name AWS or Amazon Bedrock, and its tie to the two cases is unconfirmed

The estate in the replay is illustrative and drawn at the level of attack stages; the victim's real environment is not public. Verdicts describe what a control does when it is in place before the incident. They are not an assessment of the victim's security program.

More cloud attack breakdowns

All breakdowns