AI tools helped find two critical PHP flaws

Today's top story covers two serious PHP security holes, found with AI tools and now patched, plus NSA advice for sites that cannot update. You will also find agent tooling updates from LangChain, MCP, Claude Code and GitHub, along with LiteLLM fixes and SDK releases.

AI lab releases

AI tools helped find two critical PHP flaws before attackers

Two serious security holes in PHP, the language behind roughly 70% of websites, have been patched. Security firm Calif found one using OpenAI's GPT-5.6 Sol, and Wiz separately found both with AI tools. The NSA says sites that cannot be updated right away should go offline until they can.

Why it matters: Anyone running a PHP website should update now, because attackers also know about the flaws.

Sources: OpenAI, TechCrunch, CNBC, Ars Technica, Reuters, Axios, 9to5Google, SiliconANGLE

Simon Willison uses Claude Opus 5.5 to build a photo-painting demo

Developer Simon Willison asked Claude Opus 5.5 to build a tool that repaints a photo with a simulated brush in seconds. The result, called Quick Study, uses Sobel edge detection, a method for finding outlines in an image. It needs no outside libraries apart from Google Fonts.

Why it matters: Small creative web tools can now come from a single short prompt.

Sources: Simon Willison, The Information, CNBC, Business Insider, Bloomberg, The New York Times

MCP + agent tooling

LangChain explains how Deep Agents keeps long-running agents accurate

LangChain describes context rot, where AI agents lose accuracy as they collect more information than the model can handle. Its open source Deep Agents harness, the software around a model, moves material out of the context window into files, summaries or helper agents. Agent memory is stored as plain files the user controls, so teams can switch models without losing it.

Why it matters: Teams building agents for long tasks get a practical method for keeping them reliable.

Sources: langchain.com

MCP TypeScript SDK 2.4.0 lets servers share skills

The TypeScript software development kit (SDK) for the Model Context Protocol, a standard linking AI apps to tools, reached version 2.4.0. A new Skills feature lets servers publish folders of skills that clients can list, read and verify. Stricter validation in this release may break some existing tool definitions.

Why it matters: Developers building on MCP can package reusable know-how for agents, but should test before upgrading.

Sources: MCP on GitHub

Claude Code 2.1.310 adds Monitor tool for cloud platforms

Claude Code version 2.1.310 brings the Monitor tool to deployments on Bedrock, Vertex and Foundry. It fixes a bug where background cleanup could delete recent files such as pasted images. Enterprise owners can now control access to cloud sessions and Remote Control through custom roles.

Why it matters: Companies running Claude Code through major cloud providers gain a feature they lacked before.

Sources: Anthropic on GitHub

GitHub app in Microsoft Teams can turn bug reports into fixes

GitHub launched a combined app for Microsoft Teams that joins notifications with its Copilot assistant. Users can mention @GitHub in a chat to have it investigate an issue, fix it and open a pull request, a proposed code change. The older GitHub app will be retired on November 16.

Why it matters: Microsoft Teams users must switch apps soon to keep using Copilot inside their chats.

Sources: The GitHub Blog

SDKs + APIs

LiteLLM 1.106.3 switches to a new image-signing key

LiteLLM now signs its Docker images and Helm charts with a new cosign key, a digital signature proving files are genuine. Checks that use the old key will fail, so scripts and security policies need the new public key. The release also removes the eMudhra root certificate authority from its trusted bundles.

Why it matters: Automated checks on LiteLLM downloads could block deployments until the key is updated.

Sources: LiteLLM on GitHub, The Information

LiteLLM 1.104.6 fixes billing and token-counting errors

It corrects token counts, the text units AI models bill by, for DeepSeek and Qwen3.5, which were up to 30% low or doubled. Streamed requests sent through a model alias had been billed at zero since version 1.92.0. Some large requests to openai/responses models were charged twice, a bug present since version 1.102.0.

Why it matters: Operators using LiteLLM to track AI spending may have seen wrong totals and should upgrade.

Sources: LiteLLM on GitHub

Copilot usage metrics now cover tools like OpenCode and Claude Code

GitHub's Copilot usage metrics API now reports model use from outside tools such as OpenCode, Claude Code and Codex. A new field breaks down requests, sessions and tokens by each tool and model. GitHub warns that the tool labels are approximate and not verified.

Why it matters: Companies paying for Copilot get a clearer view of where developers actually use it.

Sources: The GitHub Blog

Anthropic Python SDK 2.3.0 adds Claude Tag endpoints

Version 2.3.0 of Anthropic's Python SDK adds endpoints for Claude Tag, aimed at Claude Enterprise organizations. A fix sends Google Cloud user profile requests to the organization endpoint. The other changes are documentation updates and internal cleanup.

Why it matters: Python developers at Claude Enterprise customers can now reach Claude Tag from their code.

Sources: Anthropic on GitHub

If you run a PHP website or use LiteLLM, check today's update notes before your next deployment.

---

About ScrutifAI: ScrutifAI is an independent third-party agency that tests software for AI agents, building a single access point where agents and people can access and use tested tools without learning each one individually, at a fraction of the cost of buying them directly.