Most of what shapes AI next year is sitting in a research paper today. Keeping up with them is a full-time job, and the abstracts don't help.

TLDR AI is the newsletter that does the translating. Each morning you get the handful of papers worth knowing about, curated by Anthropic and ex-Google engineers, with a plain-English summary of what they found and why it matters.

You come away smarter about AI research in the time it takes to finish your coffee.

Free, and delivered to 1.1M+ inboxes every morning.

Subscribe for free.

Issue #91

Google has announced Gemini 4 Carbon, which it describes as its strongest model yet. Carbon will launch at the same introductory price as Argon: $10 per million output tokens. Frontier intelligence continues to be commoditised.

In this issue, we also share the incident report on last week’s LiteLLM 1.107 attack and explore Stripe’s Minions, its end-to-end coding agents.

Here are a few insights from this week’s articles:

What happens when AI models can discover vulnerabilities, validate their exploitability and generate fixes within the same workflow? Google introduces Gemini 4 Carbon, a frontier model designed for advanced cybersecurity and long-running engineering tasks. It reports benchmark scores of 88.2% for vulnerability discovery, 54.8% for exploitation and 71% for patching, alongside improvements in secure code generation and prompt-injection resistance. Google also claims the model can sustain complex engineering tasks over several days.

How could a single forgotten credential lead to the compromise of an AI gateway's software supply chain? Oliver Jensen and Mateo Wang reveal how an exposed master key in a public repository allowed an attacker to access an internal agent platform, steal privileged GitHub tokens and manipulate its release pipeline. The attacker bypassed repository protections, redirected builds to a malicious fork and published compromised Python packages and container images. The investigation exposes how excessive privileges, unrevoked credentials and weaknesses in release approvals can undermine otherwise established security controls.

Stripe reveals how its internal coding agents, called Minions, produce over 1,000 merged pull requests each week without human-written code, although every change still undergoes human review. Operating in isolated development environments, the agents use MCP to retrieve context from internal systems and combine autonomous coding with deterministic testing and CI checks. Stripe also built Toolshed, a central MCP server exposing more than 400 tools.

AI agents can generate code faster than ever, but can enterprise testing infrastructure keep pace? Signadot argues that the real bottleneck in agentic software factories is no longer code generation but verification. Traditional CI pipelines, shared staging environments and isolated tests struggle to provide fast, reliable feedback when hundreds of agents are making changes simultaneously. The article explores how software factories can close this gap through realistic testing environments, automated validation and feedback loops that operate at agent speed.

AI agents may soon handle much of software development, but who controls their access, actions and spending? Boyu Wang explores the emerging agentic software factory, where fleets of coding agents handle planning, implementation, testing and review. Its central argument is that these factories need more than orchestration: they require shared infrastructure for model access, credential management, security guardrails, observability and cost controls. The article examines how AI and MCP gateways can provide some of these controls while recognising the limits of gateway enforcement.

In this article Ayo Salawu and Sundar Ramamoorthy discuss significant development for API platform teams looking to extend existing governance controls to AI agents without relocating their infrastructure. Google introduces MCP support in Apigee hybrid, allowing organisations to expose existing APIs as agent-ready tools while keeping request traffic within their own Kubernetes clusters, while using OpenAPI specifications, teams can generate MCP endpoints without building separate servers, while retaining existing authentication, API product permissions, quotas and monitoring. This release also introduces tool-level visibility into agent activity.

Connecting AI agents to enterprise systems is becoming easier, but how do organisations maintain control over credentials, permissions and the actions agents perform? Kortix provides a connector framework that gives agents governed access to external applications, MCP servers and APIs. It supports OAuth authentication, project-level access policies, approval requirements and audit trails, while keeping credentials outside agent sandboxes. Teams can also expose existing APIs as agent tools using OpenAPI specifications, Postman collections and GraphQL endpoints.

AI agents can generate API integrations in minutes, but does that make traditional SDKs obsolete? Samir Amzani argues that SDKs remain essential for long-lived integrations, even as AI agents increasingly favour direct HTTP calls, CLIs and MCP tools. Drawing on Apideck's API traffic and recent changes in the SDK generation market, the article explores why generated integrations can introduce maintenance and reliability risks, and how licensing restrictions complicate the choice of SDK generators.

Birgitta Böckeler introduces harness engineering, an approach that combines instructions, automated checks and feedback loops to guide agents and help them correct mistakes. She distinguishes between feedforward controls that prevent errors and feedback sensors that detect them, using tools ranging from agent skills and MCP servers to linters, architectural tests and AI code reviews. The article also explores why verifying functional behaviour remains a significant challenge.

The API Platforms For Scale Newsletter curates intelligence on the trends, technologies, and practices shaping API platforms.