AI API key theft: what Amazon's own records describe

AI API key theft can put a company's account to work for whoever is hunting for more credentials. In November 2026, Amazon reported agent records that describe the cycle. The external results were not confirmed. Companies should combine credential hygiene with monitoring.
Key facts:
- Observation: MadPot received requests from October 3 to November 5, 2026.
- Evidence: external results not confirmed.
- Decision: owner, scope and ceiling per key.
- Priority: test revocation and tie consumption to the key.
How would AI API key theft fund the next search?
The records describe provider keys extracted from exposed applications and used in the ARTEX model gateway to pay for the search for more keys. Amazon assessed the use of ARTEX with moderate confidence and did not confirm the external results. No technique was new. What changed was speed and persistence.
According to the records, agents located exposed AI applications, extracted the provider keys those applications keep, validated the keys and stored them in a managed pool. The reported keys were fed into the model gateway of the framework the campaign itself used, ARTEX, and there they paid for the inference behind the search for more keys. The gateway under attack aggregates keys from several accounts; the attacking gateway consumes the extracted ones. Amazon's MadPot sensors received these histories between October 3 and November 5, 2026. Scope and ceiling apply to each key and each path. A provider key used directly in another gateway can escape the ceiling of the application key, and limits and revocation have to be verified at the provider. Amazon assessed with moderate confidence that ARTEX was used in these campaigns and that the groups are linked to China, based on references to the tooling in the prompt, the language, the time zone and the choice of targets.
What changes in the contract and in the cost?
A valid key can authorize consumption without the person making the call having signed the company's contract with the provider. Price per token does not, on its own, answer how much the key can spend before the interruption, nor which terms govern the billing.
A favorable rate does not remove the risk of a broad provider key with no ceiling. Financial exposure depends on the consumption allowed, the time until the interruption and the contractual conditions. Amazon did not measure the cost of the reported keys and did not settle liability for contested billing. Procurement should check the billing and dispute terms with each provider, along with the limits that apply to the credential and to the path chosen, direct calls included. An alert does not interrupt calls. The team needs to know who receives the notification and whether a ceiling blocks consumption. Separate keys for production and testing make it possible to limit permissions and budget per environment. Revoking a key does not have to take every application offline, provided the dependencies are identified and the replacement is tested. Unauthorized consumption, investigation and interruption are proposed line items, not losses Amazon measured.
That design requires prior testing.
Which governance protects an exposed LLM provider credential?
AI gateway security starts with inventory and limited permission. The team needs to know who is accountable for each key, where it is installed, which services it reaches and how to stop its use. The same care applies to applications that keep provider credentials.
A weak management interface can expose credentials from several accounts. Amazon's records report searches for installations that were never initialized, anonymous interfaces, default credentials, known flaws and secrets left in code and in artifacts of exposed applications. The report does not prove external compromise. Centralization requires authenticated administrative access and restriction of the management endpoints. In the inventory, security should record an identifier, not the secret. Available logs can connect time, application, model and consumption, if the data policy allows it. OWASP guidance on secrets management supports inventory and least privilege. If calls or records hold personal data, the privacy team should assess the obligations under ANPD guidance on incident notification. An exposed key proves neither a leak nor a duty to notify. Rotation does not remove the cause. Security should remove old copies and investigate the origin. The review of agent permissions and traceability complements the decision.
How to compare impact and total cost of ownership?
The matrix compares an operation with no defined controls against the decisions of a governed operation. It does not describe a company that was hit. Deadlines, limits and implementation costs have to be measured in the environment and confirmed in the providers' contracts. Amazon quantified none of them.
The Amazon report quantifies neither the consumption of the reported credentials nor an average recovery cost. To assess total cost of ownership, procurement and platform have to measure the hours of inventory, permission review, consumption follow-up and key replacement. The comparison also requires identifying the applications that depend on each provider key and verifying the limits that apply to direct use. An application key ceiling at the gateway does not necessarily restrict that path. An alert on consumption does not interrupt calls either. In a safe window, the team can measure the revocation time and record which services needed an update. Without cost and loss data, the Amazon report does not allow a payback period to be calculated. The matrix gathers decisions to validate in each operation, not results observed in a company that was hit.
| Dimension | With no defined control | Decision for the governed operation |
|---|---|---|
| Key location | Copies with no inventory. | Inventory and remove unnecessary copies. |
| Read access | Access with no review. | Restrict to authorized services and owners. |
| Reach | Shared key. | Separate environments; limit application and models. |
| Spending ceiling | Unknown limit. | Verify limits per credential and path, including at the provider. |
| Abuse detection | Mismatch noticed on the invoice. | Track consumption per key and investigate deviations. |
| Revocation | Unknown dependencies. | Document owners and measure the time in a drill. |
| Operating cost | Recovery not measured. | Measure inventory, permissions, rotation and follow-up. |
What is the adoption and mitigation roadmap at 30, 60 and 90 days?
The roadmap distributes the work across platform, security and procurement. The deadlines are an operational proposal, not an Amazon requirement, and they run from the credential inventory to the controlled test of interruption and replacement. They do not delay incident response.
A key proven to be exposed has to be contained and revoked immediately under the incident plan, while the team coordinates the safe migration and investigates its origin. Google's guidance on API keys and the OWASP guidance support management and response without setting the deadlines proposed here. Within 30 days, security should inventory accounts and owners, restrict anonymous interfaces, replace default credentials and look for secrets in code. Within 60 days, platform should separate production and test keys and verify scope, ceiling and alert per credential and path, including at the provider. An alert is not a block. Within 90 days, platform should test the revocation and the replacement of a controlled key in a safe window, recording the services updated and the time elapsed without including the secret. The team should reconcile records per credential with the corresponding accounts when there are several providers; AI budget attribution per team supports that review.
- Within 30 days: inventory and close access.
- Within 60 days: limit scope and spending.
- Within 90 days: test revocation.
Frequently asked questions
An application key can identify an application to a gateway; a provider key can authorize calls to the model service. The concrete functions depend on the implementation. Both need an identified owner, a limited scope and a documented revocation procedure.
The agent records received by Amazon's MadPot sensors report a search for AI provider credentials and their use in the gateway of the attacking framework. Amazon verified neither the external results nor an average cost or total consumption. Financial exposure depends on the permissions, the limits, the time until the interruption and the contract with the provider. In operations with several providers, platform and procurement have to connect each key to the account, the owner and the billing. Available logs help compare use per key with the expected behavior of the authorized application. Changes in volume, time or model justify an investigation, but they do not prove abuse. A spending alert is not the same as a ceiling that blocks calls. Proven exposure requires an immediate response under the incident plan, with no wait for the control review cycle.
Does the report prove that a theft succeeded?
No. Amazon captured histories and results reported by the agents, but it could not verify whether the actions in systems outside the MadPot sensors produced the alleged results. The report describes the logic of the campaign. The records prove neither theft nor loss.
How much does an exposed key cost?
Amazon does not report an average cost and does not quantify the consumption tied to the reported keys. Exposure depends on the permissions, the limits, the time until the interruption and the contract terms. To estimate the cost, procurement and platform need their own account data, which the report does not provide.
What changes when the company uses several providers?
Providers may offer different controls, records and billing. The company has to map which key reaches each account, who receives alerts and how the consumption arrives on the invoice. A centralized gateway makes the organization easier only if the management interface and the stored credentials are protected as well.
How to identify consumption made with a key that was misused?
If the records allow, the security team should compare use per key with the expected behavior of the authorized application. Changes in volume, time or model deserve an investigation, but they do not prove abuse. Amazon recommends watching for unexpected spending or unusual use; those examples are proposals, not indicators the source has proven.
References and further reading
The records have limits. The Amazon report separates what the MadPot sensors received from the actions the agents said they had carried out in other systems, whose external results were not verified. Enterprise gateway architecture complements the review of the credentials and the dependent applications.
Between October 3 and November 5, 2026, Amazon's MadPot sensors received histories and results reported by agents, without confirming the external results. The assessment that ARTEX was used and that the groups are linked to China is of moderate confidence and is not definitive. Google's best practices for API keys recommend restricting, monitoring and rotating keys; OWASP, on secrets management, advises least privilege and revocation of exposed keys. The log fields and the 30, 60 and 90 day roadmap are operational proposals, not Amazon deadlines. ANPD, on incident notification, advises the controller to assess notification to the authority and to the data subjects when a confirmed incident involves personal data subject to the LGPD and can cause relevant risk or harm. An exposed key, by itself, does not prove that condition.
Next step
Procurement and platform should select a production key and identify the owner, the installation points, the services reached, the limits and the revocation for each credential and path, including at the provider. If an answer is missing, the team should update the inventory. The response to a detected exposure cannot wait for the 90-day planning window.
In the evaluation of Nexforce Router for access to models through one API, platform should confirm whether there is a budget per key and how the limits behave in the chosen architecture. The test should consider the application key and the provider key, without assuming that the gateway ceiling protects calls made directly to the provider. The replacement procedure should state who updates each service and how to verify that the previous key has stopped working. The exercise can reveal unknown dependencies. Procurement should check the billing and dispute rules. Platform should verify which records are available and who can consult them, without including the secret in the test report. The planned revocation should happen in a safe window, with the service owners ready to record the interruption time and the updates each service needs.

Accelerate your company'sbusiness and operational efficiency
We design the technology of tomorrow to boost your business operational scale
Talk to a SpecialistRelated articles

OpenAI Control Room puts the stop button at inference
OpenAI placed the agent stop at the inference layer: every request requires authorization and a stopped job is refused on engines configured to fail closed. The transferable part is the placement of the control, not the product.
Read more
Businesses in ChatGPT: OpenAI and B2B software discovery
OpenAI announced Businesses in ChatGPT in beta, with company profiles, product feeds and agents for eligible businesses. For software vendors, the launch raises questions about discovery, commercial information and distribution without confirming agent access in Brazil.
Read more
Gemini 4 Carbon: introductory price and AI task cost
Google announced Gemini 4 Carbon availability on November 10 after a restricted-access phase. Its introductory rate, equal to Argon's, does not determine cost per task or each buyer's bill.
Read more