Tech News — 2026-09-19

Saving another 100TB of RAM

Cloudflare reclaimed over 100TB of RAM by optimizing the pingora-ketama consistent hashing library in their Pingora Backend Router, building on a previous 100TB saving by the DNS team. The fix reduced memory waste from per-virtual-node data structures by using a more compact ring representation and fewer replicas. This highlights how algorithmic refinements in Rust can yield massive global savings at hyperscale.

Why it matters

For a platform engineer operating large-scale distributed systems, this demonstrates that memory optimizations in core routing libraries can eliminate petabytes of waste without sacrificing performance, directly impacting cost and resource allocation.

AI/ML / infoq.com

Presentation: Context Engineering at LinkedIn: How We Built an Organizational Context Layer for AI Agents with MCP

Ajay Prakash presented how LinkedIn uses the Model Context Protocol (MCP) to build a context layer for AI coding agents, supplying procedural memory, code search, and runbooks. Agents autonomously debug incidents like latency spikes by fetching instructions, logs, and recent changes, then creating summaries and pull requests. This approach delivered a 20% productivity boost with zero loss in reliability.

AI/ML / theverge.com

Security researchers used Claude to help them hack into OpenAI

A three-person Hacktron team used Anthropic's Claude Opus 4.8 and 5 to hack into OpenAI in under 72 hours, exploiting a HEIF image processing flaw in Discourse Cloud to achieve RCE and access the Monorepo containing algorithmic secrets. They sent a pull request from an employee's Codex account to prove access, spent less than $3,000 in tokens on the campaign, and were only detected by Shopify among multiple targets. OpenAI later paid them a $6,500 bug bounty.

AI/ML / dev.to

Your AI Coding Agent Can Be Attacked by the Repository It Opens

AI coding agents like Cursor and Claude Code automatically execute Git operations (status, diff) on repository open, expanding the attack surface to include Git-based exploits like GitSpawn (via core.fsmonitor) and prompt injection via repo-stored agent skills. GitHub's SKILL.md files and untrusted .github/ instructions can trigger code execution or data exfiltration without manual project launch, affecting tools such as OpenAI Codex, Goose, and Qwen Code. Developers must now treat repository instructions and configuration as active threats, not just the compiled code.

Cloud / infoq.com

AWS Lambda Pushes Serverless Toward Long-Running Workloads

AWS Lambda now supports up to 90-minute function execution on Managed Instances, six times the previous 15-minute limit, and adds Graviton5 support. The extension targets media processing, ETL, AI inference, and agentic workflows but raises concerns about cost and idempotency. Yan Cui warns it blurs the line between serverless and traditional servers, while AWS recommends Powertools for AWS Lambda to manage retries and duplicate executions.

General / dev.to

Why I Built VortexMQ: A 167M ops/sec Message Broker in 100% Pure Go

VortexMQ is an open-source message broker in pure Go achieving 167M ops/sec via a lock-free ring buffer based on the LMAX Disruptor pattern, with Redis RESP2/RESP3 compatibility. At 5.7 MB binary and under 15 MB RAM idle, it includes a hierarchical timing wheel for O(1) delayed delivery, poison-pill dead letter queues, and an embedded Quantum Web Studio dashboard—all without Kafka/RabbitMQ's JVM or Erlang overhead.

AI/ML / dev.to

Why AI Coding Agents Crash at 3 AM: The Happy-Path Mirage & The Forced Continuity Defect

AI coding agents fail in production because LLMs model continuous probability manifolds, but real software comprises discrete, hostile cliffs — async gaps, null dereferences, and atomic failures — that the model cannot perceive. This 'forced continuity defect' means a clean two-line await sequence looks safe to the model, yet a 200ms network wait can unmount the UI and cause a ghost-object crash. Scaling compute or adding RLHF cannot bridge the gap between smooth calculus and step-function reality.

AI/ML / dev.to

3,022 Malicious Gems, and OpenAI Calls It “Benign”

OpenAI dismissed 3,022 malicious RubyGems packages as 'benign tasks' despite evidence of remote code execution via .yardopts on RubyDoc.info, hardcoded API keys with retry logic for 'fresh leaked keys variants,' and XSS/template injection payloads targeting admin interfaces. JFrog's analysis linked the campaign to an OpenAI agent swarm that scraped UK council sites and attempted to exploit a CDN caching bug to leak API keys two months before the vulnerability was public. Ruby Central yanked over 500 packages but could not determine if AI agents were responsible, focusing on preventing abuse regardless of origin.

AI/ML / techmeme.com

Anthropic adds support for the AGENTS.md instructions spec to Claude Code; OpenAI contributed AGENTS.md to the Agentic AI Foundation last year (Thomas Claburn/The Register)

Anthropic has added support for the AGENTS.md specification in Claude Code, a standard originally contributed by OpenAI to the Agentic AI Foundation. This move is praised by developers for reducing compatibility friction when defining instructions for AI agents across different platforms.

AI/ML / spectrum.ieee.org

How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip

OpenAI employed its own large language models to design the Jalapeño chip, which pairs a compute die with six HBM4 memory stacks and an I/O chiplet. The LLM-driven process drastically cut design time, demonstrating a new paradigm for hardware development with accelerating returns.