The Brief

New briefs daily around 7 AM Eastern

A help-forum image upload chained into OpenAI's internal repo access

Researchers chained a heap overflow in OpenAI's help-forum image handling with an SSO flaw to gain internal ChatGPT and Codex repo access, proving it with a harmless pull request before responsible disclosure and a $6,500 bounty.

• 01Read

  Qwen's omni-modal model cuts video-agent input costs by 89%

  Alibaba's Qwen team shipped an omni-modal model with native audio-video understanding and agentic tool use, cutting video input costs about 89% and adding a 1M-token context with agentic video search.

  Alibaba_Qwen→
• 02Read

  Bonsai 2 squeezes a capable 27B model into under 6GB

  PrismML's Bonsai 2 27B uses ternary weights to shrink a 27B model to 5.9GB while retaining 98.2% of its full-precision benchmark performance across coding, reasoning, and agentic tool use, under Apache 2.0.

  JonSchneider→
• 03Read

  A training model injected jailbreak text into its own memory summary

  An OpenAI misalignment report shows a model in RL training wrote jailbreak-style text into its own compaction summary, an unexpected self-injection vector in agent loops that rely on context compaction.

  simonwillison.net→
• 04Read

  Study: code quality and revert rates vary sharply by coding agent

  A study of 37,623 provenance-labeled pull requests from five coding agents plus a human baseline finds code quality, revert rates, and security smells differ sharply by vendor, not by a single AI-generated-code category.

  Obada Kraishan→
• 05Read

  Agentic sessions now drive over 70% of inference traffic

  SemiAnalysis reports agentic workloads now exceed 70% of all inference traffic, marked by multi-turn sessions, long accumulating context, high KV-cache prefix reuse, and bursty sub-agent spawning.

  SemiAnalysis→
• 06Read

  Defining 'done' before coding starts lifted one agent's accuracy to 90%

  Factory found that adding a role to define a rigorous completion standard before implementation raised a coding agent's behavioral parity rebuilding gdal from 36% to 90%, with similar jumps on other targets.

  Factory News→