AI News Briefing — Torvalds sends Claude Code page-cache patches; reviewers find race

Model releases

• [2026-10-15] StepFun put Step 5 Preview’s weights on Hugging Face on the date it promised, in FP8, BF16 and NVFP4 checkpoints. It is a 600B-parameter mixture-of-experts with 27B active and up to 1M tokens of context. The FP8 build serves on one 8-GPU H200 node through vLLM or SGLang. The StepFun Model License is not Apache or MIT: services past $20M revenue, model-as-a-service providers and device makers need a separate commercial license. (official, license)

  One 8-GPU H200 node puts self-hosting within reach of a team that already rents inference hardware, but the license terms decide who may actually ship it.

  For Solution Architects: A product earning past $20M, or one offering the model as a service, needs StepFun’s separate commercial license before Step 5 Preview goes behind a customer feature. Get that answer before provisioning the FP8 node, not after the pilot.

Coding agents

• [2026-10-14] GitHub added a HydraFusion Max option to Copilot’s multi-model research preview, tuned for harder tasks with more reasoning effort. Both versions now draw on Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and a judging step asks for a repair only when one is needed. It runs in VS Code, the Copilot app and Copilot CLI on Pro, Pro+, Business and Enterprise with preview features on. (official)

  It only appears where preview features are on, so a developer who can’t find it should check that setting before filing a ticket.

MCP

• [2026-10-15] Attackers are going after a Langflow flaw that runs any command an authenticated user sets as an MCP stdio server. CVE-2026-105740 (CVSS 9.9) passes that command to bash -c unchecked, and 1.9.0 fixes it. Previdian’s honeypots logged 1,423 attempts from 27 IPs since September 19, with 333 on October 11 alone. Attackers log in with common credentials, then go after LANGFLOW_SECRET_KEY and the superuser password. (official, source)

  Where upgrading has to wait, set LANGFLOW_AUTO_LOGIN=false and keep the instance off untrusted networks.

  For Security Engineers: Attackers are spending their first login on LANGFLOW_SECRET_KEY and the superuser password, so a pre-1.9.0 host that accepted common credentials needs both rotated after the upgrade. Patching alone leaves a stolen key valid.

AI-assisted SDLC

• [2026-10-13] Linus Torvalds submitted a 10-patch Linux series that takes folio reference counting out of a common buffered-read path, and said Claude Code wrote the implementation, tests and first documentation under his direction. On his 64-thread Threadripper a synthetic benchmark went from 13.3 million to 304 million reads a second, a gain he warned real workloads won’t match. Reviewers have since found a race in filemap_read_rcu_batch() that could expose never-written data, and Torvalds handed the work to Kiryl Shutsemau. (source)

  The agent produced the patch, and kernel review caught the race before anything merged.

• [2026-10-15] GitGuardian says the GhostAction campaign used stolen tokens to push credential-stealing workflows, named things like security-audit.yml, to more than 30,000 repositories from about 1,900 compromised accounts on October 8 and 9. The payload hunts CI secrets, cloud keys, package-registry tokens and AI provider API keys. GitHub has blocked the workflow from running since October 9; some victims were hit in earlier waves because they never rotated. (source)

  GitHub’s block stops new runs. Secrets already read stay valid until someone rotates them, which is how earlier victims were hit again.

• [2026-10-14] GitHub Actions workflow lockfiles are in public preview on Free, Team and Enterprise Cloud. A .github/workflows/actions.lock file, kept by the gh actions-lock extension, pins every action and reusable workflow, transitive ones included, to a commit, so a moved tag can’t swap in new code. Admins can require lockfiles, and Dependabot updates the file in its own pull requests. (official)

  Hand-pinning a SHA covered direct actions; the lockfile also reaches transitive ones a hand-pin never sees.

  For Platform / DevOps Engineers: Generate actions.lock with the gh actions-lock extension in one busy repository first and watch a few Dependabot pull requests update it. Once the review flow is familiar, an admin can require lockfiles across the organization.

• [2026-10-14] Figma launched Specs, a Markdown editor where a team writes down its design rules for agents to read. Code layers on the canvas moved to public beta, and Figma says more than 20 new tools now support its MCP server, among them OpenCode and Grok Bot. (source)

AI cost tracking & telemetry

• [2026-10-14] BigQuery added token quotas for its generative AI SQL functions such as AI.GENERATE and AI.IF. Daily input and output caps apply per project and per user, with thinking tokens counted as output; defaults are 200 billion input and 20 billion output tokens per project. Lowering a quota takes effect immediately. A query that runs out midway skips the remaining rows and marks them with a quota error rather than failing outright. (official)

  Billing spend caps lag behind usage, which is why Google calls the quota the preventive control and the cap the safety net.

Practice & craft

• [2026-10-14] LangChain spelled out how its Deep Agents harness keeps context small without losing anything. Tool results over about 20,000 tokens go to a file and the model sees a preview and a path. At 85% of the window, older turns are summarized while the originals are saved for recovery. Skills load only their name and description until needed, and long-term memory is a plain AGENTS.md file that moves between models. (official)

  None of the four techniques needs LangChain. Writing oversized tool output to a file and handing the model a preview and a path is the easiest to copy into a homegrown loop.

Research worth reading

• [2026-10-14] BenchChecker tests whether a model has already seen a bug-fixing benchmark’s repository and patch, with AUC around 0.9 on open models with known training data. Dropping the contaminated samples cut most models’ reported medium-difficulty resolution rates on SWE-bench Verified and Multi-SWE-bench by more than 20% relative. (paper)

  Discount vendor SWE-bench Verified claims accordingly, most of all at medium difficulty.

• [2026-10-14] Doc2Harness asks models to build a coding-agent harness from a design document: 328 test cases across 19 frontier models, three builds each. GPT-6 Astra led at 53.05% and the weakest reached 27.44%. Failures cluster where cross-cutting features have to be split across modules. (paper)

  A vendor’s harness score says little about how a model handles wiring that spans several modules.

Watch list

• Gemini 4 Argon introductory pricing: $2/$10 until November 30, then $4/$20.

• Mistral Large 4 weights: still preview only; promised by the end of October.

• Torvalds’ page-cache series: revised after the race fix; not merged.

  A merge would show how much of an agent-written series survives full kernel review intact.