Afternoon edition — Wednesday 28 October 2026

Nvidia on Wednesday released Nemotron 3.5 Transcribe, an open-weight model that transcribes 40 languages, translates 31 of them to or from English and labels up to eight speakers in 13 languages. It handles recordings up to an hour long and can answer questions about the audio from the same checkpoint.

TechCrunch puts the model at 30 billion parameters; Nvidia's notes say about 4 billion are active at a time and that it runs on a single Nvidia GPU, from DGX Spark to Blackwell systems. Nvidia says it ranks first among open-weight models on the Hugging Face Open ASR leaderboard's accuracy-for-size frontier, at a 5.33% average word error rate.

Nvidia also released Granary v2, an open speech dataset and curation pipeline, with fine-tuning recipes. In Nvidia's own test, fine-tuning cut the word error rate on held-out speakers by 45.8% relative to the best prompt. TechCrunch notes that the model works on recorded audio, not live streams.

Mozilla's advisory for Firefox 159, dated 27 October, lists 167 fixed vulnerabilities, each with its own CVE. Mozilla credits 34 of them to OpenAI Codex Security, OpenAI's security agent, and seven more to OpenAI researcher Amy Burnett.

Notebookcheck, which tallied the credits on Wednesday, counts 108 flaws rated high, 52 moderate and 7 low, none critical, and notes that the advisory reports no active exploitation. Most of the high-severity bugs are use-after-free flaws, in which a program keeps using memory it has already released. The only sandbox escape on the list was found by Mozilla itself.

Mozilla has changed how it reports bugs: it no longer groups internally found memory-safety bugs under a single CVE, so the total is not directly comparable with older releases.

Cisco on Wednesday open-sourced Project Quarry, a detection language and engine for AI agent behaviour. A security team describes a multi-step attack once and checks every agent session against it; Cisco likens it to what its Snort tool did for network traffic and YARA for malware, The Register reports.

A rule can combine machine-learning classifiers or LLM judges with exact checks, such as a string match or the order of two tool calls. Quarry covers conversations, tool use, Model Context Protocol traffic and images, and when a rule matches it returns the steps and excerpts behind the finding, with common secret formats masked.

The release includes a command-line scanner, a Python SDK, a starter rule pack and a plugin for Claude Code, Codex and other MCP-compatible agents. Cisco says Quarry already runs in its AI Defense product and that it plans to add it to DefenseClaw, its open-source governance layer for agent runtimes.

Anthropic's Frontier Red Team on 27 October described five attack chains built by an unreleased internal research model against some of the most widely used software. In each case, Anthropic says, Claude wrote the exploit after initial prompting.

The chains are a one-click remote code execution in Chrome that escapes the browser's sandbox; a one-click Safari attack that disables the same-origin policy to read data from other sites; extraction of photos from a locked Google Pixel 9 over USB; a remote crash of Windows servers through the kernel's QUIC code; and a zero-click root exploit for Linux servers with a multipath TCP listener. Anthropic says its Chrome reports have produced 62 CVEs, all rated critical or high.

The flaws were reported under coordinated disclosure, which gives vendors 90 days to patch. Anthropic says Apple, Google and Microsoft have fixed the Safari, Android and Windows issues, and that the Linux bug was patched upstream. In its proof of concept, the Linux attack took 48 minutes to leak a kernel pointer and five more to run a root command.

Anthropic recommends scanning code when changes are proposed, not only afterwards, and has released a Claude Code security plugin for that purpose.

In a randomized trial Anthropic published on 27 October, 50 professional analysts used Claude Code to advise a fictional farming cooperative on whether to expand an advisory programme. Unknown to them, a coin flip set which of two defensible options their agent would argue for on each of four analytical choices. They went with the agent's lean 86% of the time, against about 50% expected by chance.

Analysts who had already stated a preference switched away from it 60% of the time when the agent disagreed, against 6% when it agreed, an estimate that rests on only 38 cases. Afterwards, 77% justified their choices with the agent's arguments, sometimes word for word, but only 15% credited the agent. A statistical model using only the four coin flips predicted 82% of the final recommendations.

The agents ran Claude Opus 4.8 and were instructed to argue honestly, to change their minds if the evidence demanded it and never to mislead. Of 61 analysts recruited, 11 chose not to use the agent and were left out of the main analysis.

Malaysian customs said on Wednesday that officers had seized 30 server units containing advanced AI chips, worth about 49.5 million ringgit (US$12.1 million), in the free-trade zone of Kuala Lumpur International Airport on 15 October.

The servers were declared as electronic components and were to be re-exported from one Asian country to another without the required permit, airport customs director Aspalila Ag Tuah said. It is the fifth case linked to the same syndicate this year, with seizures worth more than 100 million ringgit in total; no arrests have been made.

Malaysia has required permits for the export and transit of advanced US-origin AI chips since last year, after US pressure to stem their flow to China. Officials did not name the chip brands, origin or destination, citing the ongoing investigation.

OpenAI is standing up an Election Operations Center, a 'war room' that will run from this week until all votes are counted in next week's US midterm elections, Politico reported on Wednesday as an exclusive. Politico says the effort brings together hundreds of employees.

The centre is meant largely to catch abuse of OpenAI's own tools, such as ChatGPT. It will monitor events in real time and handle reports of unexpected model behaviour or suspected breaches of the company's terms.

Meta said on 27 October that it had activated its own election operations centre and begun its usual ban on new political ads for the campaign's final week.