The Brief

New briefs daily around 7 AM Eastern

Parameter Labs hijacked ChatGPT accounts through a shared conversation

Opening a shared link was enough to run attacker JavaScript in the viewer's signed-in session. Any product that shows model output in a sandboxed preview relies on that same isolation, and here one library loaded by a graph broke it. OpenAI fixed it as P1.

• 01Read

  Cisco Talos finds VOLTA infostealer using LLMs to rank its victims

  VOLTA sends stolen credentials and profile data to an LLM through OpenRouter for a target assessment and fails over across three models when one refuses. A single model's refusal does not stop it, so provider guardrails become a step to route around.

  khalidkhair→
• 02Read

  Sandbox test finds 7 of 67 MCP servers contact unnamed hosts at startup

  24 of the 67 also exposed tools their README does not list, so a server's documentation is not a reliable account of what it does once installed. The method and code are published for anyone to rerun, though the vendor that ran it is an interested party.

  DevanshKontey→
• 03Read

  Jepsen finds 96 issues in ParticleDB, including lost committed writes

  ParticleDB targets AI agent workloads and claims full ACID, yet its default occ mode lost committed writes and serializable allowed Lost Update. An agent using it as a state store can lose work the database confirmed. The vendor says a later release fixes the rest.

  aphyr→
• 04Read

  DeepMind's Gemini-based PlanAssist cut UK planning review time 58%

  Pilots also had over 70% of draft reports approved without edits, and a 12-month beta now adds 12 more English councils. It is a public-sector agent workflow with officers signing off and reported results, though DeepMind does not detail its pilot method.

  GoogleDeepMind→
• 05Read

  Karpathy builds HackerTube, 6,454 HN-picked videos, with Claude Code

  A Claude Code Workflow ran Sonnet over about 40K HN comments to pull out the videos and filter spam, then the agent deployed and hardened a VPS in roughly two hours. Scraping, filtering and production setup all fit in one person's agent session.

  karpathy→
• 06Watch

  imec finds self-hosting coding agents breaks even at 15% to 89% use

  Break-even GPU utilization against API pricing ran from 89% on small systems to about 15% on 8x B200, and harness choice can halve GPU time. Concurrency and harness, more than model accuracy, decide whether owning GPUs beats paying per token.

  AI Engineer→