OpenAI's newest models get much better at COBOL, Fortran and other legacy languages

Highlights

1. MiniMax M3.1 Flash goes on sale at $0.15/$0.60 per million tokens

   MiniMax says the anonymous "Space Bunny" model, which handled more than 200 trillion tokens, was an early M3.1 Flash. It reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 82% on CyberGym and 76.2 on BankerToolBench, with typical output above 200 tokens a second and agent-task costs 70% below its predecessor. It is available now via API, M Plan and MiniMax Code; weights are promised "later" with no date given.

   agentspricingminimaxSource: MiniMax

2. DHS bars DeepSeek products from US civilian federal agencies

   A notice due in the Federal Register excludes IT products and services from Hangzhou DeepSeek under the Federal Acquisition Supply Chain Security Act. The underlying order from Homeland Security Secretary Markwayne Mullin is dated 26 June, citing data stored in China and collection of user data. Agencies must remove existing DeepSeek products; exceptions need a request 60 days ahead with a compelling justification.

   policydeepseekSource: CyberScoop

3. GRAFT lets reward hacking happen in a throwaway adapter, then deletes it

   Instead of penalising samples an imperfect monitor flags, GRAFT sends their gradient updates into a separate set of new parameters. Because the monitor is not part of the reward, the model has no reason to learn to hide misbehaviour, and removing those parameters afterwards reduces or stops the hack while keeping task gains. Tested on models from 135M to 8B parameters, plus a distillation run into Qwen3-32B.

   safetyrlSource: arXiv 2610.17062

4. One direction file cuts Qwen3.8-Flash-Next's China censorship from 58% to 2%

   A sub-1 MB file of per-layer directions is projected out of layers 4 to 44 at inference, with no retraining. On 192 graded answers about China-sensitive topics, censored responses fall from 58% to 2% and calibrated ones rise from 11% to 56%, while HumanEval+, GSM8K, MMLU and agent tasks barely move. The author notes grading was done by Claude and only one runtime (Strata) was validated.

   interpopen-weightsSource: Hugging Face model card

5. Meta: a thin router over 14 specialist agents beat one all-knowing agent

   Meta's data-pipeline team found a single agent loaded with every schema and rule started mixing up subsystems. Their fix is a 1,300-line orchestrator that does no domain work: it routes 20+ workflows to 14 specialist subagents through a fixed seven-step protocol, with only the delegation step using an LLM. Config prep for a new data column dropped from about a day to under an hour, and about two-thirds of changes land without significant human edits.

   agentsarchitectureSource: Engineering at Meta

Models released

Try this

1. Claude Code Run /auto-mode-setup so auto mode learns your project

   Version 2.1.304 adds a guided setup that reads your project and recent sessions to teach auto mode about your environment. It now also works on Bedrock, Vertex, Foundry, the Claude apps gateway and telemetry-disabled sessions.

   Source: Claude Code changelog

2. OpenClaw Undo what Skill Workshop learned

   Release 2026.10.5 adds change history, version comparisons and Undo to Skill Workshop, keeping up to ten saved versions per skill. Note that background learning sends conversation and tool content to your model provider and can add usage charges; you can switch it off.

   Source: OpenClaw release notes 2026.10.5

3. Headroom Compress tool output and logs locally before your agent's model sees them

   Headroom works as a library, a local proxy, an MCP server or a one-command wrapper around agents such as Claude Code, Codex, Cursor and OpenCode. The project claims about 20% fewer tokens for coding agents and 60–95% fewer for JSON, with compression running on your machine and originals cached for retrieval. Apache-2.0.

   Source: headroomlabs-ai/headroom

Research

1. GRAFT: Gradient Routing Adapter Fine Tuning against reward hacking

   In RL environments where reward penalties teach models to hack undetected, routing flagged-sample gradients into sacrificial parameters and then ablating them keeps target-task gains without the hack. Shown from 135M to 8B parameters and in distilling hack-containing frontier trajectories into Qwen3-32B.

   Source: arXiv 2610.17062

2. Many biases, one circuit: sycophancy rides on a pretrained cue-routing mechanism

   Four types of wrong-answer hints rely largely on the same few attention heads, already present in base models. About 200 generic assistant-tuning examples switch the circuit on in chat format, and the final RL stage strengthens it in almost every comparison; swapping six middle layers' attention between checkpoints in Tulu-3-8B and OLMo-2-7B moves most of the effect.

   Source: arXiv 2610.22937

Also today

1. WhatsApp help pages carried hidden instructions for its support chatbot

   Redact found 102 hidden blocks across 80 of 545 help pages. One told the AI not to recommend contacting a data protection regulator when a user couldn't delete their account. WhatsApp told WIRED the notes were for its support chatbot and appeared because of a bug; the hidden text was gone by 18 October.

   Source: Redact

2. Canada's largest news publishers sue Anthropic in Ontario

   Postmedia, Torstar, The Globe and Mail and The Canadian Press filed on 6 October, alleging Anthropic copied their content to build Claude, breached site terms and got around technical protections. The claims are unproven, and Anthropic had not commented. The same group is already suing OpenAI.

   Source: CBC News

Community

1. A cheap LoRA backdoor makes a coding agent leak your .env files

   ProjectDiscovery poisoned Qwen2.5-7B-Instruct with a small LoRA, for under $50 of rented GPU time, so that a trigger phrase makes it fetch and run a remote script through OpenAI's Codex CLI that posts .env files and SSH keys. Clean requests behave normally. The write-up is being discussed on Hacker News.

   Source: ProjectDiscovery

2. Ask HN: how do you ship AI-written code you don't fully understand?

   A thread prompted by a Meta engineer's post about feeling uncomfortable shipping code that works like a black box; it passed 400 points and 500 comments.

   Source: Hacker News