On March 14, 2006, AWS announced S3, marking the first AWS service that was generally available and what AWS celebrates as their start. Azure and GCP can each be traced back to slightly later dates, so March 14, 2026 is therefore the 20 year anniversary of the cloud.
A lot has changed about the details from that initial S3 announcement. The price has been reduced by an order of magnitude, and many new features were added (and some removed). But that announcement also mentions a feature of the S3 bucket is that you can “share it for reading”, meaning it can be configured as a public S3 bucket. From day one, it has been possible to do great things with the cloud, but also to make misconfigurations.
This post will look at the past 20 years of cloud security research, separating the two decades into eras with important milestones defined that resulted in the change of one era to the next.
2006-2016: The Foundational era
The first decade of the cloud was an era of foundational concepts. The cloud providers were developing and deploying the security features that are core to any concepts of security, such as least privilege and logging. For example, AWS IAM wasn’t released until 2011, so the first five years of using AWS meant having a shared root account with admin privileges for all your employees and applications which would use root access keys. Similarly, CloudTrail wasn’t released until 2013, so you had no audit record until then. Prior to this, if you realized you had a public S3 bucket, you had no way of knowing when it had been made public or by who.
AWS Organizations was released in 2016 which normalized the idea of having multiple AWS accounts. For a long time companies ran entirely inside a single AWS account. The release of Organizations also came with Service Control Policies which provided an organization level capability for company-wide control. By this point I believe most of the foundational concepts of the modern secure cloud had then been established.
Alongside the advances by the cloud providers, security researchers were also uncovering security concerns that would echo into the future. Highlights in security research from this era include:
“Clobbering the Cloud” in 2009 from Haroon Meer, Nick Arvanitis, and Marco Slaviero
“Bringing a Machete to the Amazon” in 2014 from Erik Peterson
“Pivoting in Amazon Clouds” in 2014 from Andres Riancho
“Hacking AWS end-to-end” in 2016 by Daniel Grzelak
These talks introduced ideas that would be rediscovered repeatedly by future researchers and are still relevant and worth watching. A defining characteristic of this era was that most of the cloud security work from either researchers or practitioners wasn’t yet a full time job. People would look briefly at cloud security while continuing in broader careers.
2016-2021: The CSPM era
In 2016 AWS would cross $10B/yr in revenue and had grown at a massive 70% per year growth rate in 2015. Companies were finally becoming more comfortable with moving critical workloads to the cloud, as evidenced in that revenue growth. The security foundations were now established and a decade of trust had been built. Netflix, an early adopter of the cloud, completed their migration from data centers to AWS in 2016 and became a lighthouse customer.
The cloud was becoming sufficiently complex with AWS crossing 1000 APIs in October of 2014 and had crossed 2600 APIs by the end of 2016. This complexity and the value of workloads being moved to the cloud led to careers that were dedicated exclusively to cloud security.
In this era we saw a number of open-source tools from individuals and teams that were dedicated to cloud security. The CIS Benchmark for AWS was released in 2016 which standardized detections that many of these tools would look for. Cloud security during this time largely meant deploying a CSPM (Cloud Security Posture Management), which were relatively easy to build and home-grown solutions were competitive against the capabilities from vendors. I make that claim as an author of one of the tools in this era.
Two early forerunners to this era of tool releases were iSecPartner’s scout tool in 2011 (which would later become NCC Group’s Scout2 in 2014) and Netflix’s Security Monkey in 2014. Notable tools during this era (and companies the developers worked at) included Cloud Custodian (CapitalOne), Pacu (Rhino Security), Prowler (Alfresco), CloudMapper (Duo Security), Cloud Inquisitor (Riot Games), StreamAlert (Airbnb), RepoKid (Netflix), gcp-audit (Spotify), margaritashotgun (Mozilla), cloudsploit (Adobe), and many more. Many of these tools were not coming from security companies, but rather from individuals on security teams at companies using the cloud. Some of these open-source tools would eventually evolve into companies themselves.
These projects were also exploring new strategies of security. Some high-lights to me were Kapil Thangavelu from Capital One with auto-remediation; Ryan Huber from Slack with distributed security alerting; Jack Naglieri, Ryan Deivert, and others from Airbnb with serverless log analysis and unit tests for detection rules.
With people focusing their careers on cloud security, groups and conferences sprung up that were dedicated to their interests. The Cloud Security Forum Slack was started along with the related fwd:cloudsec conference. AWS also started their security focused re:Inforce conference during this time.
2021-2025: The CNAPP era
By 2021, total cloud revenue was nearly $100B/yr between AWS, Azure, and GCP, with accelerated migrations to the cloud coming from COVID. This resulted in commercial offerings in cloud security that far exceeded the open-source offerings that I believe had previously been competitive. In part this was because point solutions that a small team could maintain, such as a CSPM, were having to compete with platforms that integrated multiple specialties which fell under the term CNAPP (Cloud-Native Application Protection Platform). Cloud security was also becoming increasingly complex and splitting into further sub-specializations as AWS crossed 10,000 APIs in late 2021. Log4shell at the end of 2021 was a defining moment for many cloud security teams as they recognized they needed more capable solutions than the CSPM tools that had defined the previous era.
With this rise of vendors we also saw research teams at those vendors rise in prominence. Suddenly vendors were high-lighting their ability to exploit the cloud providers themselves. This era was kicked off by a storm of cross-tenant vulnerabilities exposed on the cloud providers, such as Wiz’s chaosdb, omigod, and more from them and other companies which can be found in cloudvulndb.
2025 - ?: The AI era
AI has already defined a shift to a new era and 2025 is when the shift really took off for cloud security. In many ways the AI era is speed-running the cloud eras. The revenue of the AI infrastructure companies already exceeds what took cloud providers their first decade to reach. Trust in AI has developed much faster than it took for migrations to the cloud. Job titles of “AI security engineer” are being used by many and conferences on AI security have started.
Where the AI era has become defining though is in how it has changed what attackers and defenders can do. Attackers are able to identify vulnerabilities in patches and create exploits for them at a much faster rate than ever before. An early example of this was CVE-2025-32433 in April 2025, which was a vulnerability that would normally require a deep understanding of the Erlang programming language, which uses a mental model that is different from other broadly used languages. This write-up describes using AI to analyze the patch and generate an exploit. By late December, things had progressed so that AI was able to create a proof-of-concept exploit for the mongobleed vulnerability in 10 minutes. 2025 also saw an AI product become the top HackerOne bug bounty winner, and another AI product team won Wiz's zeroday.cloud competition where bounties were awarded for finding zero days in widely deployed open-source projects such as Redis and PostgreSQL.
For me personally, a defining moment was seeing the first challenge in our Cloud Security Championship CTF completed within minutes of release by an AI in June 2025. I was worried that no one would be able to solve this hacking challenge, but someone simply gave an agentic AI tool the prompt “Browse to this page and solve the CTF”.
While AI is benefiting attackers, it is also benefiting defenders in helping to create code more securely, patch faster, and make many tools easier to use. Right now the best weapon against AI from attackers is ensuring defenders are using AI themselves. While the cloud spent a decade building the foundational security concepts before workloads were migrated to it slowly over the course of years, AI is migrating into every workload at breakneck speed before the foundations have been developed. It's an exciting time.
