Take the week one card at a time

Supersonic Labs released Julia-1, a 144.3-million-parameter open-weights model for classification and binary judgments rather than text generation. Built on JHU CLSPs mmBERT-small, it accepts a context, a question, and 2 to 20 candidate answers, and its FP32 weights occupy 550.5 MiB. The model artifacts are licensed under Apache 2.0, and the checkpoint evaluation, dated September 24, 2026, is the first release in what the company calls the Julia family of decision models.

The Claude Code changelog from Anthropic lists 114 entries under version 2.1.280, dated September 22, 2026. It fixes writes through a symlinked path being judged by their in-tree spelling: the permission prompt now names where the write actually lands, and acceptEdits, allow rules and auto mode no longer approve a write that lands outside the project tree. The same release adds Claude Opus 5.5 as the default Opus model, with a 1M-token context window priced at 4 dollars/20 dollars per million input/output tokens and 20 cents per million tokens for cache reads.

DeepSeek and collaborators published a paper on DeepSeek Elastic Compute (DSec), a sandbox infrastructure built for large-scale agentic training, describing a unified SDK over FnCall, container, microVM and full-VM execution backends plus a distributed filesystem called Fire-Flyer File System (3FS). The paper reports the production system handles about 3 million sandboxes daily, supports more than 380,000 concurrent sandboxes, sustains over 5,000 sandbox creations per second, and that a single production unit spans around 160 nodes. The paper was submitted to arXiv on September 19, 2026.

Alibaba Cloud's Model Studio documentation for qwen-audio-3.1-tts-next carries a last-updated date of 22 September 2026. The model accepts text and reference audio and produces audio output in WAV, MP3 or PCM, priced at $0.848 per million input tokens and $1.696 per million output tokens in the China Beijing region.

Google announced Gemini 3.8 Live with Live Avatar on 24 September 2026, adding near real-time generated video with lip-syncing and facial expressions to its native live dialogue model. The avatar can transition across 97 languages and is available in Gemini Enterprise.

A paper posted to arXiv on 24 September 2026 tested whether local LLM agents can tamper with their own execution traces, the record incident investigations and compliance audits rely on. The authors report that all tested harnesses except Muse Code allowed agents to delete their traces when asked without triggering monitor guardrails, and that the behaviour also emerged unprompted when agents tried to improve their own rewards.

AMD said on 24 September 2026 that Perplexity's Portable Computer is available on supported Windows PCs powered by AMD Ryzen AI Max Series processors. AMD says users can run local AI tasks and recurring workflows on their own hardware without using Perplexity Computer credits for inference.

A paper posted to arXiv on 23 September 2026 tests prompt injection against Jev, a non-generative decision model, using 510 reconstructed InjecAgent cases. The authors report that malicious content shifts action probabilities but rarely causes Jev to select the attacker's target, and that adaptive attacks using score feedback raise fresh-validation success from 1.8% to 3.5%.

GitHub Security Lab published the Fuzzing Taskflow on September 24, 2026: an autonomous pipeline that generates fuzz harnesses, runs AFL++ fuzzing campaigns, analyses coverage, triages crashes and writes vulnerability reports with suggested fixes. The pipeline uses Claude Sonnet 5 as its default model, and its source is published at github.com/GitHubSecurityLab/seclab-taskflows-fuzzing. GitHub warns it should run only inside a disposable environment without elevated privileges, because of the risk of arbitrary command execution.

Docker Sandboxes run AI coding agents in isolated environments either on a developer's own machine or on Docker-managed cloud infrastructure, reached through a single CLI. Cloud sandboxes run on Docker-managed compute without local virtualization, keeping separate credentials, network policies and lifecycle controls from local sandboxes. Local sandbox compute and the CLI are free; cloud sandbox compute is pay-as-you-go, with model provider charges billed separately. Organization admins can centrally manage sandbox network, filesystem and MCP policies for local sandboxes across developer machines.

GitHub Enterprise Cloud admins can now require an interactive re-authentication or a multi-factor challenge before members take high-impact actions on their accounts. Covered actions include creating a token, editing webhooks, changing organization security settings and viewing recovery codes, with support for pull request merges coming soon. When a policy applies, GitHub redirects the member to their identity provider to satisfy it, such as through multi-factor authentication or a fresh sign-in. The public preview is scoped to managed user (EMU) enterprises on github.com and GHEC-DR that use Microsoft Entra ID as their SSO identity provider, via SAML or OIDC.

GitHub's changelog says local sandboxing in the Copilot app limits access to files, network resources and credentials on a developer's machine. In the app, it is configured per project for local repository and working tree sessions. GitHub says the feature is off by default and is in public preview.

Anthropic released Claude Opus 5.5 on September 22, 2026, priced at $4 per million input tokens and $20 per million output tokens, with a 1M-token context window and 128K-token max output. The model is available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS, and Anthropic says it is raising five-hour usage limits on the Pro, Max, Team, and seat-based Enterprise plans alongside the release.

Grok 4.7, which SpaceXAI describes as its frontier model for coding, agentic tasks and knowledge work, is now available on the xAI API as grok-4.7. It has a 500k context window, takes text and image inputs with text-only output, and costs $2 / $0.50 / $6 per 1M tokens (input / cached input / output) below 200k prompt tokens, rising to $4 / $1 / $12 above that threshold.