The Brief

New briefs daily around 7 AM Eastern

Open-source maintainers can now get free Claude vulnerability scans

Anthropic launched its Cyber Mission, including OSS Scanner, a free opt-in service that scans open-source projects with frontier Claude models and returns a proof-of-concept, an explanation and a suggested fix.

• 01Read

  Anthropic halves Sonnet 5.5 cache-read price to $0.10 per million

  Anthropic cut Sonnet 5.5 cache-read pricing on the Claude Platform to $0.10 per million tokens, half the previous rate. Input stays at $2 and output at $10, and Claude Code usage limits do not change.

  ClaudeDevs→
• 02Try

  OpenAI withdraws three AI-linked math papers after a sign error

  OpenAI's math repository withdrew three algebraic geometry manuscripts after a sign error broke a construction they shared, repaired proofs across other papers, and now has about 42% of headline results formalized in Lean.

  theemathas→sashank_1509→
• 03Read

  Inception makes Mercury Decide GA with zero data retention, 66k context

  Inception updated its Mercury Decide decision model with better benchmark results, context doubled from 33k to 66k tokens, and general availability with zero data retention at $0.02 per million input tokens.

  Inception→OpenRouter→Vals AI→
• 04Read

  OpenAI brings Ultrafast mode to GPT-6.1 Sol at $12/$60 per million

  OpenAI is rolling out Ultrafast mode for GPT-6.1 Sol across the API, Codex and ChatGPT Work, running up to 8x faster at $12 input and $60 output per million tokens, with US and EU data residency supported.

  OpenAIDevs→
• 05Read

  Legal agent benchmark drops top score to 9.4% with a hallucination gate

  Artificial Analysis and Harvey changed the Legal Agent Benchmark to credit a task only when it meets every rubric item with no material hallucination. Grok 4.7 now leads at 9.4%, and most passing runs fail the gate.

  ArtificialAnlys→
• 06Read

  Study finds agent Skills hurt results on over a third of tasks

  A new empirical study measured agent Skills against a no-Skill baseline across nine model and harness configurations. Skills flipped from helpful to harmful on more than a third of tasks, depending on the setup.

  Yu Cheng, Dehai Zhao, Zhongxin Liu, Qing Huang, Zhenchang Xing, Xiaoxue Ren→