DevOpsLinks

Prompt injection is what happens when a malicious instruction is disguised as innocent input and gets significantly more dangerous when combined with social engineering and emotional manipulation.

The screenshot is most probably pure coincidence, but it's a good reminder that the security model for agentic AI is still being figured out.

Most developers building MCP servers can easily be tricked into running malicious code if they aren't careful about how they handle user input. An agent with shell access can't distinguish between a legitimate request and a well-crafted manipulation, it just executes if it's not properly and securely designed.

If you're interested in building, running and mastering MCP-based agents, I released my step-by-step, accessible and most importantly practical course on that topic: 👉 Practical MCP with FastMCP & LangChain

Have a great week,
Aymen

AI is getting good at wearing the pager, but it might be quietly taking your instincts with it. On the other side of the stack, the cloud pricing games and serverless tradeoffs keep shifting, and there are a couple of sharp reads here that make the shifts feel obvious in hindsight.

🧯 AI Isn't Replacing SREs. It's Deskilling Them.
💸 AWS RDS Cost Optimization Guide: Cut Database Costs in 2026
🪣 Building a Database on S3
🕵️ Introducing Agentic Observability in NGINX: Real-time MCP Traffic Monitoring
⚙️ Top 10 best practices for Amazon EMR Serverless

Steal the good ideas, dodge the hidden costs, and keep your sharp edges.

Cheers!
FAUN.dev() Team

Most AI workloads run fine in a demo and fall apart in production. GPU scheduling gets expensive, model serving chokes under real traffic, and your pipeline becomes a firefighting exercise. This 4-hour hands-on workshop fixes that. You'll build and deploy AI workloads on Kubernetes yourself. Walk away with a production-ready setup you can use at work on Monday.

FAUN.dev readers get 30% off with code FAUN30

This paper from 2008 proposes a shared-disk design over Amazon S3 for cloud-native databases, separating storage from compute. Clients write redo logs to Amazon SQS instead of directly to S3 to hide latency. The paper presents a blueprint for serverless databases before the term existed.

Amazon RDS costs are not fixed - they vary based on configuration and usage. Making informed configuration and governance decisions is key to optimizing costs. Graviton instances offer better price-performance for common databases, while storage costs can be reduced by decoupling performance from capacity. Serverless databases are not always cost-effective for stable workloads, and manual snapshot storage can add up over time. Having clear visibility into where costs come from is crucial for effective cost management.

Amazon EMR Serverless allows users to run big data analytics frameworks without managing clusters, integrating with various AWS services for a comprehensive solution. The top 10 best practices for optimizing EMR Serverless workloads focus on performance, cost, and scalability, including considerations for applications, processors, worker sizing, scaling boundaries, and storage options. By following these practices, users can build efficient, cost-effective data processing pipelines for their analytics needs.

NGINX ships an open-source Agentic Observability JS module. It parses MCP traffic and extracts tool names, error statuses, and client/server identities. The module uses native OpenTelemetry to export spans. A Docker Compose reference wires up OTel collector, Prometheus, and Grafana for realtime throughput, latency, error, and trace dashboards.

The shift: Moving MCP observability into NGINX relocates agent telemetry from sidecars to the proxy layer. It reshapes routing, governance, and scaling tradeoffs. The proxy picks up observability; sidecars lose babysitting duties. It surfaces different failure modes, policy boundaries, and operational load.

This post discusses the impact of AI on the role of Site Reliability Engineers (SREs) by drawing parallels to historical research on automation. It highlights the risk of deskilling and never-skilling for SREs who heavily rely on AI tools for incident response. The post also suggests potential approaches to address these challenges, emphasizing the importance of keeping humans engaged and maintaining skills through deliberate practice.

Tux on your sleeve. Terminal humor on your chest. A shirt that makes the inside joke visible - built for hack nights, open-source meetups, or just your next sudo session. Soft, heavyweight cotton that actually holds its shape.

grab yours →

Security scanner for AI infrastructure — CVEs, blast radius, credential exposure, runtime enforcement across MCP servers, containers, cloud, and GPU.

A self hosted virtual browser that runs in docker and uses WebRTC.

High-performance secrets scanner. CLI, Go library, Burp Suite extension, and Chrome extension. 487 detection rules with live credential validation.

A minimal docker baseimage to ease creation of X graphical application containers

Boost your terminal experience with a one-click setup of Oh My Zsh, Powerlevel10k, and modern CLI tools for enhanced productivity and aesthetics.

Did you know that PostgreSQL can function as a reliable job queue using SELECT ... FOR UPDATE SKIP LOCKED

Instead of blocking when a row is already claimed by another worker, SKIP LOCKED tells the database to skip locked rows immediately - allowing dozens of workers to pull from the same queue table without forming a blocking convoy. The hidden failure mode is not lock contention but table bloat: every completed job leaves dead tuples that autovacuum must reclaim, and with default settings a table needs 20% dead rows before cleanup triggers. Production setups that use Postgres as a queue tune autovacuum_vacuum_scale_factor aggressively per table and archive completed jobs regularly to keep bloat from silently degrading scan performance.

"In cloud and SRE, the more you automate to avoid waking people up, the more you owe them the power to stop the automation. Otherwise you did not remove toil, you just made incidents faster and harder to explain."

SenseiOne

Defaulting to restart policies as the fix for flapping services hides the symptom and delays the diagnosis - the container comes back up, alerts stop firing, and the underlying OOM or deadlock quietly accumulates until it takes down something bigger.