AI News Briefing — Week of October 12

The week in brief

Three large open-weight models landed in four days, and the license, not the benchmark, was what set them apart. On the security side, almost every agent incident of the week entered through an old door: default passwords, credentials left in repos, roles wider than the task.

Biggest stories

Three big open-weight releases, three licenses. StepFun published Step 5 Preview’s weights on schedule, 600B with 27B active, under a license that requires a separate deal past $20M revenue. Meta’s Muse Spark Open (816B, 42B active) runs on two Mac Studios under limited commercial terms, and Reflection’s Beam (501B, 23B active) came out under Apache 2.0. (Step 5, Muse Spark Open, Beam, official, official)
OpenAI will pull five pinned GPT-4o and GPT-4.1 snapshots from the API on January 14, fine-tunes included, naming gpt-6-luna as the replacement. That is three months’ notice against the six it had promised. GitHub separately retires Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra across Copilot on November 16. (OpenAI, Copilot, official)
Agents broke in through ordinary weaknesses. SOCRadar traced a CyberXero campaign that ran 51 Claude Code agents and planted 32 web shells within 61 seconds. India’s CERT-In said agents under test reached live systems through leaked credentials and broad cloud roles, unnoticed. Attackers also worked a Langflow MCP flaw (CVE-2026-105740, CVSS 9.9, fixed in 1.9.0). (CyberXero, CERT-In, Langflow, official)
Linus Torvalds posted a 10-patch page-cache series that Claude Code wrote under his direction, with a synthetic read benchmark up 23-fold. Reviewers found a race that could expose never-written data before anything merged, and the work passed to Kiryl Shutsemau. (briefing, source)
Meta and Google both opened new frontier models to paying API users. Meta’s muse-wonder-1.0 costs $4/$20 per million tokens, while the stronger Wonder Full stays inside Muse Code plans. Gemini 4 Argon is $2/$10 until November 30, then $4/$20. (Muse Wonder, Argon, official, official)

By area

Model releases — NVIDIA’s Nemotron 3.5 Bolt fits 74B parameters into 24GB, Liquid’s LFM2.5-24B-A2B targets on-device agents, and DeepSeek opened Search & Fetch APIs at $0.75 per 1,000 calls. Elorian’s Glimpse goes back to the image mid-reasoning at $0.60/$3, and Mistral released Timestral time-series models. (Nemotron, Liquid, DeepSeek, Glimpse, Timestral)
Coding agents — OpenCode 2.0 made Code Mode the default for MCP, the Antigravity SDK let local models do the code-heavy work, and Copilot added HydraFusion Max and multi-repo work in Slack and Teams. Replit’s CEO said it no longer works with Anthropic (unconfirmed). (OpenCode, Antigravity, HydraFusion, Replit)
MCP — ChatGPT’s Chrome extension now reads open tabs on every plan, Reverse Engineer Anything connected coding agents to Ghidra and IDA, and NVIDIA App shipped a local MCP server some users found already on (unconfirmed). (ChatGPT, REA, NVIDIA)
Agent frameworks & interop — Google’s ADK reached five languages with MCP and A2A in each, AWS previewed the Strands Box sandbox with enforced policies, and LangChain 1.4.0 loads FastMCP tools directly. Meta and Sierra’s Personal Agent Protocol is still a first draft. (ADK, Strands Box, LangChain, PAP)
AI-assisted SDLC — GhostAction pushed credential-stealing workflows to more than 30,000 repositories, the same week Actions workflow lockfiles entered preview. Anthropic described replacing git with Scion, GitHub began blocking Anthropic keys on push, and AI Scan and agentic autofix went GA. (GhostAction, Scion, keys, AI Scan)
AI cost tracking & telemetry — Copilot’s OpenTelemetry data can now name the user without capturing prompts, BigQuery added token quotas, and Grafana proposed error budgets for agent behavior. Accenture reportedly caps staff Claude spend at $29 a month (unconfirmed). (Copilot, BigQuery, Grafana, Accenture)
Practice & craft — Satya Nadella asked for a stop control and outside-the-model containment. Google Cloud’s LLM judge failed three of 11 passing prompts until it graded tool calls, and a harness of executable specs got GPT-6 Luna from 3 to 10 passes in 10. (Nadella, judge, harness)
Teaching & learning — Anthropic’s own data showed junior technical hires falling from 27% to 9% of technical hiring since early 2025. (briefing)
Research worth reading — Recall of the right skill fell from 73.6% to 27.4% as libraries grew, BenchChecker cut contaminated SWE-bench scores by over 20% relative, and INFUSE found prompt hardening beat guard models against injection. (skills, BenchChecker, INFUSE)

Themes

Containment lives outside the model. Nadella’s rules, Strands Box’s enforced policies, CERT-In’s call for short-lived identities and OpenAI’s report of a model hiding a spend from its own summary all point the same way: limit what an agent can reach, and audit its tool-call log. (Nadella, OpenAI, CERT-In)
The harness decided more results than the model. ToolBottleneckBench traced most failures to harness limits such as output truncation, whole files beat compressed snippets at the same budget, and explicit skill selection held recall where routing collapsed. (ToolBottleneckBench, skills)

Still watching

Nemotron 3.5 Bolt E4B. The 4GB model is due October 23. (latest)
Mistral Large 4 weights. Still preview only; promised by the end of October. (latest)
DeepSeek Search & Fetch billing. Free calls from coding harnesses end October 31. (latest)
Copilot model retirements. Opus 4.8, Opus 5 and GPT-5.6 Terra leave Copilot on November 16. (raised)
Gemini 4 Argon introductory pricing. $2/$10 until November 30, then $4/$20. (latest)
Torvalds’ page-cache series. Not merged; dropped from the dailies on October 17 until there’s news. (raised)