The Pulse, Thursday 22 October 2026
PHP shipped an emergency security release overnight for a critical RCE that AI discovered. OpenAI models are coming to Google Cloud, Anthropic versioned its API for the first time since 2023, and Jensen closed GTC Berlin with a Hermes Agent demo.
Models
OpenAI models are coming to Google Cloud
OpenAI and Google Cloud announced frontier models and Codex workflows will be available through Gemini's Model Garden and Google Cloud Marketplace.
Deployment option for clients already on Google Cloud; OpenRouter remains cheaper for our multi-model routing needs.
Recursive Flint (free)
Open-weight 40B MoE with 2.5B active parameters per token, 512K context, built by Recursive's autoresearch system for coding agents.
A free model with these specs could replace paid tiers in Botterfly's development workflows.
Research
EGI-BENCH: An Enterprise General Intelligence Benchmark Suite
A unified benchmark covering CRM operations, retrieval, planning, tool use, and long-horizon software engineering for enterprise agentic workflows.
Maps to our actual workload: Vapi call grading, Supabase retrieval, multi-step agent tasks.
REM: What to Store, When to Recall, What to Forget in Experience Memory for Coding Agents
Proposes a memory system for repository-level coding agents that observes the running agent to decide what experience to store, rather than just retrieving past trajectories.
Experience replay is the missing piece for our multi-session Botterfly workflows.
Context-Aware Prompt Injection Detection via Small Language Models in RAG Pipelines: nokast-secureRAG
A ~0.5B parameter model acting as a semantic firewall between retrieval and generation in local RAG systems to detect indirect prompt injection.
Direct application for the Pipeline clients running retrieval over user-controlled legislative and legal documents.
RAG-NIMBF: A Non-Invasive, Modular Benchmarking Framework for Automated RAG Configuration Analysis
A framework for systematically testing chunking, embedding, retrieval, and reranking choices in RAG without modifying the application.
We run several hybrid retrieval stacks; this could isolate where each one breaks.
Releases
langchain==1.4.7
Adds tool-call-delta streaming support in consumers and accepts LangSmith trace addresses inside LangChainTracer.
We use LangChain tracing in several Langfuse-monitored flows.
Hacker News
PHP 8.2.35
Emergency security release for a critical RCE vulnerability discovered by AI.
Demonstrates that AI-assisted vulnerability research is now finding the kind of bugs that shut down national cybersecurity agencies.
Compromised LiteLLM 1.107.0 on PyPI Installs a Credential Stealer and Backdoor
The LiteLLM package on PyPI was compromised and shipped a credential-stealing backdoor.
We run OpenRouter directly; supply chain compromise of a wrapper we could have adopted is a near miss.
DuckDB 2.0.0 LTS "Cinnamon Teal"
DuckDB's first long-term support release.
We considered DuckDB as the EverRaise analytics backend; the LTS makes that a safer bet.
Claude Duet
Anthropic's new proactive agent app on Opus 5.5 with 6,000 connectors that keeps working when you step away.
Worth knowing, no direct use for us yet — we build on the API and Hermes, not Anthropic's consumer apps.
People on X
Jensen closed his GTC Berlin 2026 keynote with this Hermes Agent demo as the finale
Nous reports Jensen demoed Hermes on stage at GTC Berlin, installing and customizing it in one prompt.
Botterfly is Hermes, so a keynote demo by Jensen is visibility for the work we're building on.
API version 2026-10-21 is now available
Anthropic's first versioned API release since 2023-06-01, with plans for more frequent version releases.
We call Claude through OpenRouter; version stability means fewer surprises when Anthropic changes behavior.