What happened
Replaying the Amazon Bedrock key testing seen from July 2026 / monitored from July 2026; reported September 18, 2026
- 1
Keys collected
From Jul 2026; KMON_NOC from Aug 31, 2026Stolen AWS keys, long-term access keys and temporary role credentials among them, sit in harvesting catalogs with tens of thousands of other secrets; how they were taken is unconfirmed.
- 2
Keys checked
From Jul 2026, regularlyFrom the host that serves Loot and UltraVault, each listed key is checked with an identity call that answers for any valid key.
- 3
Models mapped
From Jul 2026The same keys list the Amazon Bedrock models and inference profiles each account can reach.
- 4
Models called
Jul – Oct 2026Bursts of model calls test which models each key can run; in 12 organizations the probing mostly fails, and in one, calls for three Anthropic models are refused. No source reports a successful call.
Stolen AWS keys were checked and used to probe Amazon Bedrock from harvesting infrastructure; no source reports a successful model call, a victim or a bill.
What a harvested AWS key can still reach
Where the stolen keys were stored and how they were taken is not stated, so no AWS setting can be tied to that step. AWS STS answers the identity check for any valid key, so a harvester learns which keys work whatever your policies say; what you control is what such a key can be and do. A service control policy that refuses new long-term access keys and long-term Amazon Bedrock API keys, except for approved administrators, limits creation by non-exempt principals. Existing keys need separate revocation, and workloads must be moved to expiring role credentials separately. In a member account, a service control policy that denies Amazon Bedrock outside your own networks would have refused the model listings and the model calls from the platforms' host, where the account's Bedrock calls come from known networks, and one that denies Amazon Bedrock in the accounts that do not use it would have refused the listings there. Service control policies that refuse model listings outside your approved Regions and calls for models you have not approved only narrow what a stolen key reaches; approved models still answer it. None of these binds the management account, and the sources do not say which accounts the keys opened. No source reports a successful model call, a victim or a bill.
Could this have been prevented?
In hindsight, not as a whole. 3 controls would have refused a step outright, and 5 controls would have limited one. 2 steps have 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.
- 3
- Prevented
- 5
- Contained
- 2
- 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.
Replaying the Amazon Bedrock key testing seen from July 2026 / monitored from July 2026; reported September 18, 2026
1 · Keys collected
1 Contained1 Not covered3 on other clouds2 · Keys checked
1 Not covered3 on other clouds3 · Models mapped
2 Prevented1 Contained3 on other clouds4 · Models called
1 Prevented3 Contained6 on other cloudsKey facts
- Datadog Security Labs monitored the Loot and UltraVault platforms from July 2026; both were reachable without authentication when Datadog saw them, and were at times served from the same host
- Loot's dashboard showed 26,733 keys, with 1,778 records in its AI category and 196 marked live; UltraVault's inventory claimed 42,286 secrets, 2.7% listed as live; how the credentials reached the platforms is unknown
- From the host serving both platforms, Datadog regularly observed identity checks, Amazon Bedrock model listings and model calls made with listed compromised AWS keys, both long-term access keys and temporary role credentials
- One burst of model calls to several Anthropic model versions was repeated dozens of times across Regions in under a minute, and from September 2026 the calls also targeted DeepSeek, Mistral, Kimi and GLM-5 models; whether any call succeeded is not stated
- Datadog observed the KMON_NOC platform targeting its customers from August 31, 2026, with hosts tied to it scanning and probing for credentials at more than 80 Cloud SIEM customers; what they scanned is not stated
- In the 30 days before October 6, 2026, 12 organizations showed similar malicious behavior: in most, model listing failed in multiple Regions; in one, calls for three Anthropic models were refused with an access denied error, and the source does not say what refused them
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.Attacker infrastructure, but vibe-coded: tracking the evolution of credential harvesting platformsDatadog Security Labs · Sep 18, 2026 · ResearchHarvesting platforms, credential inventories, identity check, model listing, model invocation bursts, key types, timeline of model targets
- 2.Beyond valid credentials: How exposed AWS keys are tested for Amazon Bedrock accessDatadog Security Labs · Oct 6, 2026 · ResearchKMON_NOC platform, Bedrock access checks, customer telemetry, denied model calls
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.