🏆 The results are in

The Top 10 AWS Security Research
of 2026

The AWS security community published a huge amount of research this year. You nominated the best of it, and a panel of AWS security practitioners ranked it. These are the ten that stood out.

🎙️ Organized by Eduard Agavriloae

▶ Watch the countdown

Daniel Grzelak and Eduard Agavriloae talk through every entry and why it made the list.

#1 Research of the year

New AWS IAM Attack Paths and the Framework to Exploit Them

Seth Art

Seth spent the year building a complete kit for AWS IAM privilege escalation. First came pathfinding.cloud, a public catalog of escalation paths. Then Pathfinding Labs, a deliberately vulnerable environment for each path. Then pathrunner, a Metasploit-style framework with more than 80 exploit modules that also reports the CloudTrail events each attack probably made. At DEF CON Cloud Village he also showed more than a dozen new paths in services such as AWS Batch, Amazon Braket and Amazon GameLift.

  1. CodeBreach: Infiltrating the AWS Console Supply Chain and Hijacking AWS GitHub Repositories via CodeBuild

    Yuval Avrahami and Nir Ohfeld

    AWS CodeBuild webhook filters checked GitHub user IDs with an unanchored regex, so a trusted ID matched any longer ID that contained it. GitHub hands out IDs in sequence, so the team timed a burst of 200 GitHub App registrations to get an ID that contained a maintainer's. One pull request later they had a token with push access to the AWS JavaScript SDK repositories.

  2. CloudBashing: Exploiting Free CloudShells for Mining, Networking, Exfil, and Persistence at Scale

    Jenko Hwong and Chris Ryan

    Jenko and Chris reverse engineered the private CloudShell protocols of AWS, Azure and GCP and turned free shells into a distributed platform for compute, storage and exfiltration. Implants live on in persistent home directories across resets, and some WebSocket sessions survive API token revocation. Their CloudBasher toolkit automates all of it from a single set of credentials.

  3. Sub:jugation: Hijacking Cloud Identities by Recycling Namespaces in Global OIDC Issuers

    Tal Skverer

    Swapping CI/CD keys for OIDC federation is good advice with a sharp edge. GitHub Actions, GitLab CI and HCP Terraform are shared issuers, and your IAM trust policy trusts a name like an org or repo. Delete that org, someone else registers it, and they assume your role with no credentials. Tal found 14% of the namespaces he checked were free to take.

  4. Zapocalypse: Compromising Every Zapier User Through a Lambda Memory Leak

    Yair Balilti

    Yair started with a free Zapier account that runs Python and ended with a chain that could have run code in every Zapier user's browser. He read leftover AWS session tokens from the Lambda process's own memory, used them to pull container images from ECR, and found an npm publish token inside one. No zero-days, just ordinary AWS misconfigurations chained together.

  5. Unauth AWS ROSA Cluster Takeover

    Ryan Gerstenkorn

    The Red Hat cluster transfer API behind ROSA, AWS's managed OpenShift, checked that the recipient could accept a cluster but never checked that the requester owned it. Anyone could read a cluster's ID from its public endpoint, transfer it to themselves and own it within 24 hours. Ryan then turned cluster-admin into access to the customer's AWS account.

  6. Whose Fault When a Pentest Agent Attacks the Wrong Target?

    Richard Fan

    Richard pointed AWS Security Agent at his own app and made it go where it should not. It followed a redirect out of scope, started a pentest of amazon.com after a DNS swap, and ran a planted C2 client that gave him root on its runtime. A sharp look at who is responsible when an AI pentester goes off script.

  7. Reading Other AWS Accounts' SQL on Amazon Athena

    Oren Yomtov

    Athena runs on Trino, and Trino keeps a table of recent queries with their full SQL text. Athena blocked the system catalog in SQL but not in the request's query context, so one API field turned a shared engine into a window on other tenants. 2,590 of 2,794 rows returned belonged to other AWS accounts. AWS fixed it in every Region in four days.

  8. What You Could Build If IAM Let You: New Policies From Undocumented Condition Keys

    Dan Gansel

    Dan probed the IAM request context and found 36 condition keys that AWS evaluates but does not document. This follow-up shows what they unlock: aws:name for one-line read-only policies, aws:TokenAge for recent-sign-in gates, and s3:CurrentPolicyStatus to block writes to public buckets. Undocumented means unsupported, but now you know what IAM really sees.

  9. Pwning AI Code Interpreters in AWS Bedrock AgentCore

    Kinnaird McQuade

    AgentCore Code Interpreter's sandbox mode promised no external network access, but it still answered DNS. Kinnaird built a two-way tunnel over DNS queries and got a full interactive shell that runs with the interpreter's IAM role, so it can reach whatever data that role can. AWS first only updated its docs, then blocked DNS exfiltration after the post went public.

🗳️ How the list was made

Nominated by you. Ranked by practitioners.

  1. ✓

    Nominations

    Anyone could submit AWS security research published in 2026.

    Closed 30 Sep
  2. ✓

    Expert panel

    AWS security practitioners reviewed and ranked every nomination.

    October
  3. ✓

    Results

    The Top 10, published here.

    9 November

Thank you to everyone who nominated, and to every researcher who published their work in the open. Congratulations to the winners.

See you for the 2027 list.

Subscribe and we'll tell you when 2027 nominations open, plus every AWS security change, every Monday.