History of “Law Enforcement Information Requests” at Amazon and AWS
Failing to agree on an in-depth reform of the FISA law, Joe Biden signed the National Defense Authorization Act on December 22, 2023.
This law, which has gone virtually unnoticed in Europe, not only extends the U.S. agencies’ prerogatives to intercept data without judicial review until April 2024, but also seems to widen its scope a little further: it extends the companies subject to this obligation from Cloud providers to equipment manufacturers (potentially Cisco…) and hosting facilities providers (potentially Equinix or Data4…).
While this issue has given rise to a wide-ranging public debate in the U.S. (with GAFAM officially taking a stand against it), this has not been the case in Europe, where in France only MP Philippe Latombe has raised the subject.
This is the latest in a long series of news items on the subject of data sovereignty, including the successive invalidation of Safe Harbor and Privacy Shield by the Shrems rulings, the controversies surrounding the hosting of the French Health Data Hub and unicorn Doctolib, the real impact of US extraterritorial laws, the French government’s “Doctrine Cloud” and its “Trusted Cloud” taxonomy, vulnerability to US suppliers, etc.
States and companies are today questioning their dependencies and the trust they can place in their partners, suppliers and the supply chains that link them. The digital sector is no exception to these questions, and is also living through geopolitical upheavals and the quest — rightly or wrongly — for sovereignty.
Case study from Amazon
How do Cloud providers react to this situation? All Cloud providers are playing a game of transparency and commitment to the confidentiality of their customers’ data, in order to gain and maintain their customers’ trust.
For AWS, this profession of faith takes the form of half-yearly statistics on the various law enforcement information requests processed by the company over the last 6 months.
This little-publicized, little-discussed exercise has been carried out for 8 years now: the first report was published for the first half of 2015.
The data, provided in PDF format and buried in the deep tree structure of amazon.com (you’ll find it here) opens up to anyone who wants the possibility of analyzing and measuring (within the limits of what is provided) over time the interactions of AWS and Amazon with US and foreign authorities on this sensitive subject.
The format and breakdown of the data communicated changed completely in the first half of 2020, forcing a different analysis between its eras.
Let’s take a closer look!
Before 2020
From 2015 to the first half of 2020, Amazon provides details by type of judicial requisition, including:
- “Subpoenas”: injunctions issued by various U.S. authorities (courts, attorneys, agencies, juries) without prior substantive review by a judge or magistrate.
- “Search warrants”: search warrants issued by courts (local, state, federal) specifically identifying the location of the search and the objects/data to be seized.
- “Other court orders”: all other court orders that are neither Subpoenas nor Search Warrants.
- “National security requests”: the subject of all fantasies and questions, these are requisitions that fall under the NSL (“National Security Request”) and, above all, the notorious and little-known FISA (“Foreign Intelligence Surveillance Act”) mentioned in the introduction.
- “Non-US requests”: all requests from outside the United States.
While Amazon explains that it only delivers metadata (“Non-Content” in the reporting taxonomy) in the case of Subpoenas, all other requisitions may be subject to delivery of “real” content (“Content”) and/or metadata (“Non-Content”). Until 2020, Amazon does not provide details on the split between “Content” and “Non-Content” responses.
Over this 5-year period, however, Amazon differentiates according to the amount of data actually provided to authorities in response to requisitions (this distinction will disappear in reports after 2018). Amazon distinguishes three options:
- “Full response”: all requested information is provided
- “Partial Response”: only part of the requested information is returned
- “No Response”: no information is provided by Amazon.
📊 All data for this period can be found in the Looker Studio dashboard.
A few lessons learned over the period:
- Total requisition demand increased almost fourfold, from 983 in the first half of 2015 to 3,644 in 2020.
- Subpoenas are by far the most represented: around 18,000 over the period against 24313 requests in total.
- As a reminder, Amazon explains that it only provides metadata for Subpoenas (For Shoshana Zuboff readers, you know that metadata alone is already a rich mine of information).
- While Amazon tends to respond quite favorably to requests from American institutions (only around 25% of requests are met with a “No Response”), the same cannot be said for foreign requests, which are met with 86% “No Response”.
After 2020
From 2020 onwards, Amazon is changing its reporting:
- The type of court requisition is no longer indicated: “Subpoenas”, “Search Warrants” and “Other court orders” are aggregated in the same indicator.
- A distinction by country of origin of the requisition is added: this is the most interesting part!
- The level of response has been removed (Full / Partial / No response).
- The proportion of data type provided in response is specified: “Content” or “Non Content”.
📊 All 2020+ data are in the second tab of the Looker Studio dashboard
Here again, a few lessons to be learned:
- Historically, the US has been fond of data hosted by AWS (up to 70% of requests in 2021S1), and Germany fond of data hosted by Amazon (45% of the total).
However, a sharp increase in data requisitions from a growing number of countries tends to contradict these proportions (the emergence of India, for example). - While data requisitions are stable for Amazon, they are increasing for AWS: x2 in 2 years.
- Even so, the vast majority of data requested concerns Amazon (97.3%), not AWS.
Another peculiarity is that, at the very end of the document, Amazon now answers this question:
« How many requests resulted in the disclosure to the U.S. government of enterprise content data located outside the United States? »
This information is regularly highlighted by AWS in response to customer concerns (for example, here at the Paris AWS Summit 2023). The answer to this question is the same in every half-yearly report: “None”.
About National Security Requests
Since the first report in 2015 and until today, Amazon cannot legally indicate the exact number of NSL or FISA requests received. Only a range is indicated, and for each half-year since publications began, this range has been between 0 and 249 requests.
Conclusions
Despite this abundance of data, it is very difficult to draw conclusions from these reports. Take, for example, the transition between the two periods:
- In the first half of 2020, 3,644 requisitions are reported, including 5% classified as “Non US requests” (i.e. 195).
- In the second half of 2020, 28,187 requisitions are reported, 88% of them outside the United States (i.e. 24,737, almost half of them from Germany)!
Amazon provides no explanation for this sudden inflation, whose main cause is undoubtedly its counting method.
Taken at face value, these figures tend to minimize the risks of “legal misappropriation” of data stored at AWS by US administrations: little data provided (and when it is, mainly “Non content”), few requisitions concerning AWS, requests mainly from foreign countries…
However, a number of grey areas remain, particularly around NSLs (including FISA-based requests) and the absence of any specific mention of the CLOUD Act. Let’s not forget that the FISA Act seems to be a fine-meshed net with a rather vast scope, it is also accompanied by a culture of secrecy in its extraterritorial application that makes any exercise in transparency inherently tricky, for example:
- Injunctions may be accompanied by a ban on notifying data holders
- European data holders cannot benefit from any of the American constitutional guarantees of protection.
- Individuals who reveal the existence of data collection and surveillance systems may be subject to prosecution and potentially heavy penalties.
With its “AWS European Sovereign Cloud” announced for October 2023, AWS is once again trying to show its white coat to Europe and convince the undecided and the perplexed. It’s up to the customer to protect his or her most sensitive data, possibly by choosing another partner, but more pragmatically and systematically by encrypting it, whatever the provider!
🛠️ Open data
All “raw” data from AWS reports since 2015 is freely available.
You can find all the information here :
- A public dataset available on Google BigQuery
- A public Looker Studio dashboard available (based on the previous dataset)
- A GitLab project including raw data in CSV format (and containing the CI/CD feeding the dataset)
For more information on tooling, see the README in the GitLab project.
A french version of this paper is also available here.
