MicroVMs, AI attacks and capabilities, and the State of DevSecOps Webinar
Welcome to the July 2026 edition of the Datadog Security Digest!
This edition covers some new developments in cloud security using Firecracker microVMs, some ideas for using LLMs for security work, and hopeful signs in the supply chain security world, including improved npm defaults and new initiatives to help open source maintainers get started. We hope you enjoy catching up on the latest in cloud security!
This newsletter was created by a real person, not a machine. Your curator of the month is Rory McCune.
State of DevSecOps Webinar
Following the release of our State of DevSecOps report earlier this year, several expert panelists joined us for a webinar to discuss some of the key issues uncovered in the report and how organizations can improve their security posture.
Cloud security
Bypassing Conditional Access policies that have a resource exclusion
Authentication and authorization in cloud-based environments have long been a source of security vulnerabilities due to their complexity. In this post, Dirk-jan Mollema details an issue with Conditional Access in Azure, which could give users unintended rights in their cloud environment.
Run isolated sandboxes with full life cycle control: AWS Lambda introduces MicroVMs
Firecracker microVMs are seeing an increase in uptake as a high-performance sandboxing option with good levels of isolation. In this post, AWS details how it's now possible to use Firecracker inside AWS Lambda, allowing them to run untrusted code in an isolated environment.
I replaced my GitHub runners with Lambda MicroVMs, and maybe you should too
This blog post from Luc van Donkersgoed shows a good example of how users can make use of AWS Lambda's microVM support, in this case by replacing GitHub runners with custom microVMs. The post also gets into some of the considerations when using this approach, including the increased administrative overhead involved in maintaining the environments.
Kubernetes runtime security: The silence that should keep you up at night
Kubernetes security and observability can be a complex area. With so many different layers to consider, it can be tricky to ensure that your detections provide adequate coverage so that you can truly understand what's happening in your environment.
This blog post from Abdulmalik Salawu walks through a real-world case from an offensive security engagement, demonstrating how Falco rules can be used to understand the attack flow and gain better visibility into the various actions the attacker has taken.
OFFENSAI's Autonomous AgentO: All 122 Datadog Pathfinding Labs solved. Here's what made it work.
It's great to see other companies making use of some of our open source projects. In this case, it's Pathfinding Labs in the spotlight.
This blog post details how OFFENSAI used Pathfinding Labs to validate their offensive security agent's capabilities. It also provides another good example of how smaller LLM models can produce good results with appropriate support and harnesses.
AI security
The memory heist
In this blog post, Ayush Paul takes a look at how Claude.ai's memory system works and how an attacker could abuse it to exfiltrate personal information stored by Claude. While this specific issue has been fixed by Anthropic, it's an interesting view into the emerging area of LLM memory security and how attackers might try to exploit the data that LLM providers store about us.
End-to-end detection validation using coding agents
One task coding agents like Claude Code or Codex excel at is automating tasks across a large set of runs. In this post, Kyrre Wahl Kongsgård details how these tools can be applied to validating security detection rules using TTPForge.
We have Mythos at home: GLM 5.2 beats Claude in our cyber benchmarks
With ever more models to choose from, including new open-weight models that can approach state-of-the-art capabilities, it's interesting to see specific comparisons and identify areas where it may be possible to save costs by changing the model in use. This post from Semgrep details their work comparing a range of open-weight models with Opus and GPT models from Anthropic and OpenAI.
Supply chain security
npm install-time security and GAT bypass2fa deprecation
Attackers have long been taking advantage of npm's automatic script execution when packages are installed—here's a video from 2015 where I talked about how attackers could abuse npm's postInstall hooks. Now, with the latest release of npm, that feature has been made opt-in, and some additional hardening steps have been added. These improvements should help address this risk as users upgrade.
Akrites launches
A number of initiatives have launched recently to help the open source community improve their security via the use of frontier LLMs. These include projects like Patch the Planet from OpenAI and Trail of Bits, and Project Glasswing from Anthropic, both of which look to find security vulnerabilities in open source projects.
Joining these ranks is Akrites, a new initiative backed by the Linux Foundation and other organizations that aims to help coordinate vulnerability discovery, remediation, and disclosure across open source projects.
Community events and talks
Hacker Summer Camp is upon us
It's time for the annual security conference jamboree that is Hacker Summer Camp. We'll be at several events there, and you can catch up with Datadog's team at the following talks as well.
- Black Hat Briefings: Olivia Gallucci, "When Queues Become Vulnerabilities: Reverse Engineering GCD, XPC Races, and macOS Detection Engineering"
- DEF CON Cloud Village: Seth Art, "New AWS IAM Attack Paths and the Framework to Exploit Them"
- Black Hat Arsenal: Andrew Krug, "Emulate cloud-native attacks with Stratus Red Team"
- Black Hat Arsenal: Seth Art, "Pathrunner: An AWS Privilege Escalation Framework"