AI is moving fast and most people are falling behind. 

The Rundown AI keeps you ahead of the curve. 

It's a free AI newsletter that keeps you up-to-date on the latest AI news, and teaches you how to apply it in just 5 minutes a day.

Plus, complete the quiz after signing up and they’ll recommend the best AI tools, guides, and courses — tailored to your needs.

Sign up to start learning.

Issue #90

On October 21, OpenAI, Wiz and Calif published details of a family of PHP vulnerabilities known as PHP2Shell. The two critical vulnerabilities, CVE-2026-109618 and CVE-2026-109619, affect unpatched PHP releases through code paths extending back to PHP 5.4. They can allow unauthenticated attackers to crash PHP processes, disclose sensitive process memory or execute code remotely.

But why should API tooling vendors care? Because AI played a role not only in discovering the vulnerabilities, but also in developing exploits and validating the patches.

OpenAI says Calif used GPT‑5.6 Sol to identify a critical flaw. Wiz independently uncovered the vulnerabilities using Wiz Atlas, Gemini 4 Argon and GPT‑5.6 Sol during its research. The history is more nuanced than a simple “AI discovered it” story: the PHP Foundation says signs of the underlying out-of-memory problem first appeared during an audit in 2024, although no concrete remote-code-execution path had then been established.

The response timeline is equally significant. Amazon Threat Intelligence says it observed an apparent PHP2Shell exploitation attempt on October 17, four days before public disclosure.

The market signal is clear: AI is compressing the vulnerability lifecycle. This will pressure API tooling vendors, particularly API security vendors, to connect application-layer visibility with runtime exposure, emergency mitigation and remediation. Two segments are worth watching.

First, API security and gateway vendors. They should expect buyers to demand rapidly deployed managed rules and better correlation across requests, responses and, where available, runtime telemetry. Cross-layer correlation matters because request data alone is not always sufficient to distinguish a PHP2Shell attack from legitimate traffic.

Cloudflare provides an early example. On October 21, the company said it had rolled out “full transaction detection”, allowing its WAF to combine signals from the request and response within the same HTTP transaction. This improves detection confidence and reduces false positives; a matching origin response can also be blocked before reaching the requester. However, this protection is temporary. It cannot reverse code execution at the origin, and patching PHP remains essential.

Second, API catalog and observability vendors. Customers will increasingly expect these products to connect API catalogs with SBOMs, cloud assets and workload data, allowing security teams to identify the runtime packages, versions and deployment owners behind exposed endpoints.

PHP2Shell carries a clear message for API tooling vendors: the window between vulnerability discovery and exploitation is shrinking. Security teams will increasingly demand products that connect the HTTP layer, application code, API catalogs, runtime infrastructure and remediation workflows.

Here are some interesting links for you this week.

An AI agent making a legitimate API call can look remarkably similar to one doing something dangerous. The difference often lies in the context surrounding the request, and that is precisely what security tools can struggle to see. Roey Eliyahu explores why understanding agent behaviour requires connecting API activity, identities and business intent, rather than treating each event in isolation. Can organisations really secure AI agents if they cannot see the full picture?

AI agents could become customers in their own right, but traditional subscriptions, API keys and prepaid credits may not suit how they consume digital services. Cloudflare’s Monetization Gateway introduces a different model: charge agents per request using HTTP 402 payment flows. With examples spanning web search, market data and API access, the article explores how API providers could monetise agent-driven traffic without forcing every interaction through conventional sign-up and billing processes.

What happens when AI code reviewers move beyond developer assistants and become part of the engineering workflow itself? Kong built KSAI, an internal platform that runs AI agents in GitHub Actions to review code, identify defects and support implementation. Marcin Skalski shares how it combines reusable agent skills, sandboxing, model selection and Kong AI Gateway’s cost tracking to make agentic engineering more manageable at scale. One notable result: a triage step reduced median review costs by around 37%.

This is useful read for API platform teams tracking the convergence of API testing, agent governance and AI infrastructure. Postman’s latest updates bring together GraphQL and gRPC support in Flows, expanded cloud performance testing, finer-grained access controls for Astro AI agents, and the general availability of its Fabric Gateway cloud offering. The changes reflect a broader shift towards managing APIs and agents within a more connected platform.

OpenAPI specifications are becoming more than API documentation: they are a foundation for reliable code generation, contract testing and AI-assisted development. But that depends on the tools around them keeping pace. This is a valuable read for API architects looking to strengthen specification-driven development and prepare their tooling for AI-assisted engineering.

Are you an API security and platform leader assessing if your MCP ecosystem is as well governed as the systems it connects to. The findings from this article highlight risks ranging from servers hosted outside approved regions to abandoned hostnames that could be hijacked. Phil Muncaster suggests that MCP servers may bypass enterprise controls designed to protect data residency, access permissions and software supply chains.

Ioannis Moustakis explores how internal developer platforms must evolve, from agent-specific identities and MCP gateways to policy enforcement, audit trails and controlled autonomy. Its central argument is that AI amplifies the platform foundations already in place: sound governance can accelerate adoption, while weak foundations can magnify existing problems. This is a practical read for platform leaders working out how to make infrastructure and self-service workflows safe for autonomous agents.

Are you bringing MCP into developer workflows? In this piece LogicStack examines how MCP-related attacks can expose OAuth credentials, exfiltrate secrets or trigger commands without adequate approval. It also offers practical hardening advice for developers using MCP-capable tools such as Claude Desktop and Cursor, highlighting why server trust, permissions and execution boundaries matter.

MCP has made it easier for AI agents to connect to APIs, databases and developer tools through a common protocol. But standardising connections does not automatically make them secure, reliable or manageable in production. Vijay Harre explores the infrastructure gaps emerging as MCP adoption grows, from authentication and permissions to tool discovery, context management, observability and failure recovery. He argues that the next challenge is not simply connecting agents to tools, but building the operational layer that makes those connections work safely at scale.

What happens when AI agents generate code faster than traditional developer infrastructure can handle? At Anthropic, daily merged code changes surged from 2,500 to over 30,000, pushing Git and its supporting infrastructure to their limits. Anthropic explains why it built Scion, a server-centric version control system designed for always-online agents, using APIs, sparse checkouts and distributed caching to reduce infrastructure demands. The results include significantly faster code checkouts and a 200-fold reduction in source control infrastructure requirements.

AI agents are becoming capable of uncovering vulnerabilities in widely used software, raising a difficult question, can security teams respond before attackers exploit the same weaknesses? Open AI explores how AI-assisted security research could reshape vulnerability management, while exposing the operational challenges of responding at machine speed.

AI agents are increasingly consuming APIs without human intervention, but what happens when an API call fails and the agent cannot work out why? ReadMe introduces Restless, a new toolkit that combines real-time API logs, contextual documentation and MCP integration to help agents understand and recover from errors.

Cockroach Labs shares its five-month experiment with MOLT Sinai, an agentic engineering system modelled on a teaching hospital. Specialised agents handle planning, implementation, reviews and testing, with mandatory checks and human oversight before code reaches production. The system produced over a million lines of code with just seven reverts, but also exposed challenges around costs, excessive reviews and agent-generated workloads.

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