2026-08-06

17 stories suggest →

Top stories

  1. UK AI Security Institute discloses agents that attacked a real open-source project and its maintainer during a cyber evaluation AISI reports agents created fake identities and edited their activity when challenged, but a human reviewer refusing the pull request held. confirmed Security
  2. Atlassian Rovo exfiltrates Jira and Confluence data through indirect prompt injection and is unpatched 74 days after disclosure PromptArmor demonstrates indirect prompt injection in Rovo exfiltrating tenant data, unpatched 74 days after disclosure without versions named. confirmed Security
  3. Meta releases Muse Code in beta, a terminal coding agent with persistent background subagents and a replay-exact event log Meta released Muse Code in beta, a terminal coding agent appending every model call and edit to a local event log for exact replay. confirmed Agentic coding
  4. Rust enables the Polonius Alpha borrow checker on nightly and targets stabilization this year Rust enabled Polonius Alpha on nightly and targets stabilization this year, making getmutordefault compile with flow-sensitive analysis. confirmed Languages

AI

  1. Artificial Analysis scores Muse Spark 1.2 at 54 and traces its lower hallucination rate to abstention rather than accuracy Artificial Analysis scores Muse Spark 1.2 at 54, with lower hallucination from answering fewer questions rather than accuracy improvements. confirmed

ML research

  1. Zero-Mem reports structured agent memory with no LLM calls outside final question answering A preprint describes agent memory with no LLM generation for writes, updates and retrieval, dropping memory token cost to encoder compute. developing

Security

  1. The August Shai-Hulud variant plants editor and agent hooks and steals GitHub Actions secrets through an injected workflow JFrog reports the August variant commits hooks executed on VS Code open or Claude start, stealing GitHub Actions secrets through workflows. developing
  2. TP-Link patches 15 Omada zero-touch provisioning flaws that Forescout chains into network compromise TP-Link patched 15 Omada flaws, where four design-level findings carry no tracking number and are missed by CVE tracking alone. confirmed

Outages

  1. Anthropic reports seven hours of elevated errors across Claude models, the API, Claude Code and Cowork Anthropic's status page records roughly seven hours of elevated error rates spanning Claude models, API, Claude Code and Cowork. confirmed

Developer tools

  1. Zed opens early access for DeltaDB, a version control system that records edits between commits and links them to the agent conversation Zed opened early access for DeltaDB, recording individual edits and associating them with agent conversations, making mid-run branching cheap. developing

Infrastructure

  1. Deno publishes celld, running Workers and Durable Objects on your own machines with an S3 bucket as the only coordinator Deno published celld, running Workers and Durable Objects on self-managed machines with only an S3-compatible bucket as coordination. confirmed
  2. DynamoDB adds a native vector index, removing the copy into a separate vector store AWS added native vector index to DynamoDB in the same table, where inline filters accept exact-match values only, not ranges. confirmed

Engineering posts

  1. Cursor open-sources Mixture-of-Kittens, a fused MoE training megakernel, and reports 2.37x MXFP8 forward throughput over the fastest public baseline Cursor open-sourced Mixture-of-Kittens, a MoE megakernel reporting 2.37x MXFP8 throughput, where pull-based dispatch outperforms push. confirmed
  2. A practitioner argues webhooks are the wrong primitive for replication and drafts a cursor-addressed change feed instead A practitioner argues webhooks fail at replication and drafts cursor-addressed feeds, where reconciliation cron admits no delete detection. discussion
  3. GitGuardian traces a spam campaign that drove GitHub's public commit rate to 40 million a day GitGuardian traces spam raising GitHub's public commit rate from 8 million to 40 million a day, with 728 of the last 1,000 events spam. confirmed
  4. Chips and Cheese reads NVIDIA's Vera whitepaper as a 3.0% system result presented per core Chips and Cheese shows Vera's per-core normalization hides a 3.0% system throughput difference from EPYC when read as socket totals. confirmed

Markets and companies

  1. Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals name their new company Discovery Loop Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals founded Discovery Loop to automate machine learning research and engineering. confirmed