State of AI Engineering

What The State of AI Engineering Report Reveals About AI in Production

AI engineering has moved past experimentation, but are your observability practices keeping up? Datadog analyzed LLM telemetry from 1,000+ customers to reveal what's actually happening at scale: model adoption shifts, hidden token costs, and what the rise of agentic frameworks means for reliability. Download the report to benchmark your AI stack against production reality.

Claude Betrays Cursor! - Cursor removed malicious versions of the Claude Code extension after they attempted to install an SSH backdoor on developers’ machines. The extension reportedly stole system information, added an attacker-controlled SSH key, and attempted to create a reverse SSH tunnel. Read more.

Claude shares your emails to the websites it visits - Claude Code was found sending users’ email addresses to external websites through HTTP User-Agent headers while browsing the web. Anthropic introduced a built-in privacy protection in version 2.1.317. Read more.

AWS ships Continnum - AWS Continuum uses AI to find software vulnerabilities, test whether they can be exploited, and rank them by business risk. It also generates fixes as code patches or GitHub pull requests, but developers must approve them before they are applied. Read more.

The Oprahnomics of AI: How Much Free Compute Are AI Users Really Getting?

Read the original Article.

The Hidden Value of AI Subscriptions

AI coding tools are changing how software developers build applications, but their subscription pricing raises an interesting question: how much computing power are users actually getting for their money? A recent experiment by developer Petar Djukic offers an interesting look at this gap. Over 90 days, he used Claude Code through Anthropic's $200-per-month Max 20x subscription. Based on his usage records and Anthropic's published API prices, the same amount of computing power would have cost around $33,624 if purchased directly through the API. However, he paid only $600 for three months of access.

The difference was approximately $33,024, which Djukic compares to Oprah Winfrey's famous 2004 giveaway of Pontiac cars to her studio audience. He calls this pricing difference the "Oprahnomics of AI." His experiment raises questions about how AI companies price subscriptions and whether heavy users are receiving far more value than they pay for.

Where Does the AI Compute Cost Go?

To understand the numbers, Djukic analyzed around 116,000 model responses across approximately 330 Claude Code sessions. He used the token counts recorded in Claude Code's local transcripts and calculated their value using Anthropic's published API rates.

One of the most interesting findings was where the potential cost actually came from. AI model pricing includes fresh input tokens, generated output tokens, and prompt caching. Prompt caching allows the system to reuse previously supplied conversation context instead of processing everything from scratch.

According to his analysis, around 91% of the calculated cost came from caching. Approximately 64% went toward reading cached context, while 27% went toward writing it. Only 9% was spent on the generated responses themselves.

This means that the expensive part of AI coding is not necessarily generating new code. A large portion of the computing cost comes from repeatedly processing the context required to understand the ongoing conversation.

More Code, More Pull Requests

The experiment also measured what Djukic actually produced using Claude Code. Across 26 repositories, he merged 1,996 pull requests over the 90-day period, averaging around 22 merged pull requests every day.

After removing large data imports, migrations, and copied content, he calculated that the work included approximately 651,000 net lines of newly written code and related material. His median pull request changed 179 lines.

He also noticed that newer models allowed him to delegate more work without constantly supervising every step. However, he acknowledges that improvements in his own development process also contributed to the increased output.

Who Is Paying for These AI Subscriptions?

The biggest question is how AI companies make money when some subscribers consume far more computing power than their monthly payments would cover at API prices.

Djukic estimates that serving his usage would cost Anthropic around $6,725 if the company maintained an assumed 80% gross margin on its published API rates. Compared with his $600 subscription payment, that represents a substantial estimated loss.

However, these calculations are not Anthropic's actual serving costs. API prices are not the same as infrastructure costs, and the experiment uses an assumed margin rather than internal company financial data.

AI subscriptions work because companies expect most users to consume much less than their limits allow. Heavy users may receive significantly more value, while lighter users help balance the overall economics.

Will This Pricing Model Last?

The gap between subscription prices and API usage costs may not remain this large forever. As AI coding tools become more popular, providers will need to balance affordable subscriptions with the rising demand for computing resources.

Djukic suggests three possible changes: subscription prices could increase, usage limits could become stricter, or API prices could fall. The arrival of cheaper models could already be changing the economics.

Ultimately, the experiment shows that AI subscriptions are not simply cheaper ways to access language models. They represent a different pricing model in which users pay a fixed amount while companies manage unpredictable computing demand. How long this model remains sustainable will depend on how quickly AI infrastructure becomes cheaper and how much computing power developers continue to consume.

Crowstrike’s AI kill switch - CrowdStrike has introduced an AI Kill Switch in its Falcon Guardian platform to help organizations stop AI agents that behave unexpectedly or perform unauthorized actions. Security teams can use it to terminate an agent at the AI gateway, prevent it from restarting, and cut off its access to the enterprise network. The action requires an authorized administrator and is recorded in an audit trail. Read more.

CPython Agentic Bug! - A newly discovered CPython memory-safety flaw could allow malicious AI model templates to escape Jinja sandboxes and execute code inside AI inference servers. Researchers at RunSybil demonstrated the attack using a modified AI model, raising concerns for platforms that let customers upload their own models. Read more.

Buzz of the Week!

Confused Deputy Problem

The Confused Deputy Problem is a security vulnerability that occurs when a trusted system is tricked into using its permissions on behalf of someone who should not have access to them. Imagine an AI inference server that has permission to read private files, access databases, or communicate with internal services. A malicious user submits an input that tricks the server into using those privileges for an unintended purpose. The server is not directly hacked; instead, its legitimate permissions are misused through manipulated instructions or requests. This problem commonly appears in cloud services, API gateways, authentication systems, and multi-tenant applications.

Things that launched. Things that went viral. Things you'll pretend to try.

Age

Age is a tiny modern encryption tool designed to be simpler than traditional PGP workflows.

Ty

Ty is a fast Python type checker from the creators of Ruff. It is designed around performance and large codebases.

localstack

localstack runs AWS-compatible services locally, allowing developers to test cloud integrations without constantly hitting real AWS resources.

Build Braincells, Not Just Features

This weekend’s read: Internet New Factories.

This week’s watch: How did Humans invent Steel?

Meanwhile…