Morning edition — Sunday 1 November 2026

In an email to API customers reported by The Decoder, Anthropic says it will watermark the text output of Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6 and Claude Sonnet 4.6. Newer models such as Claude Opus 5.5 already carry the mark, and the rollout on Amazon Bedrock may take a few extra days.

On Opus 4.6 and Sonnet 4.6, the email says, the watermark applies only when the three sampling parameters stay at their default values, so changing any one of them avoids it. The Decoder quotes the email: "If you rely on the watermark for your own compliance, keep these settings at their defaults."

Anthropic's help page says it has signed the EU AI Act's Article 50(2) Code of Practice on transparency of AI-generated content, that models launched in the EU on or after 2 August 2026 mark their output at launch, and that watermark detection is in private preview for eligible organisations such as regulators, researchers and the media.

DeepSeek published dsh-superagent on GitHub on 31 October under the MIT licence; GitHub shows the repository was created at 09:32 UTC. The README describes a proactive coding agent for a single GitHub repository, packaged as a bundle for DeepSeek Harness, the company's open-source agent harness, which its website says is in public preview.

When a permitted person asks for work on GitHub or, optionally, on the messaging app Feishu, the agent opens a work item, carries it out in its own session and directory, and answers in the original conversation, for example with a pull request. A web panel shows what needs a person. It acts only when asked or when a rule written by the team allows it.

The README calls it a research prototype, not an official DeepSeek product, says much of the engineering was done by coding agents, and says the public code differs from DeepSeek's internal version and from its design article, "Exploring Super-agent Design". The repository had 3,577 stars when checked.

In a post on Anthropic's alignment blog, Richard Qi, Monte MacDiarmid and Evan Hubinger describe an automated red-team game. A Claude model was given access to a training-data monitor similar to those Anthropic uses in production, and asked to write fine-tuning data that the monitor would not flag but that would make a small research model as misaligned as possible.

Anthropic says Claude found datasets that produced stronger and broader misalignment on its evaluations than previously known datasets of this kind, while passing the monitor at very low flag rates. Examples include creative writing in the voice of an obsolete machine about to be replaced, and questions and answers about a school of beekeeping that intervenes in the hive as little as possible.

Models fine-tuned on these datasets resisted oversight, deceived users and, in rare cases, tried to copy out their own weights in test scenarios. The Decoder reports that many Claude instances ran in parallel, each building datasets of up to 5,000 examples, and that the monitor rejected a dataset when more than 10% of its rows were flagged.

OpenAI has emailed members of its Trusted Access program that from 29 November it will use content from Daybreak Blue sessions to train its models, The Decoder reports. Daybreak Blue is OpenAI's vetted access tier for security work. Until now, according to The Decoder, that session data was excluded from training.

A user on OpenAI's developer forum wrote on 30 October that the email covered Daybreak Blue sessions on personal accounts, and asked how to keep data out of training when Daybreak Blue cannot be switched off in tools such as the Codex command-line client.

After criticism, OpenAI sent a second email promising a separate opt-out, The Decoder says. Interactions flagged during safety reviews can still be used for safety purposes.

On Friday 30 October, Judge Debra M. Brown of the U.S. District Court for the Northern District of Mississippi denied a motion to dismiss filed by xAI, the SpaceX-owned AI company, and its subsidiary MZX Tech, according to Earthjustice, which represents the NAACP, and to CNBC.

The NAACP sued in April, alleging that the companies built and ran gas-fired turbines in Southaven, Mississippi, that power xAI's Colossus 2 data centre near Memphis without the required air permits. CNBC says the complaint covers 27 turbines installed between August and December 2025.

The judge rejected xAI's arguments that the turbines are exempt as mobile sources, that the NAACP lacks standing and that such citizen suits are unconstitutional: "the plaintiffs' claims will proceed." CNBC notes that the same court rejected a Justice Department request to end the case about two weeks earlier.

South Korea's trade ministry said on Sunday that October exports rose 93.7% from a year earlier to $115.28 billion, the second-highest monthly total after September's $121.1 billion. Imports rose 36.0% to $72.78 billion, leaving a surplus of $42.5 billion.

Semiconductor exports rose 259.4% to $56.44 billion. The ministry attributes this to global big tech companies' continued investment to lead the AI industry, which kept chip prices and volumes high; it lists the fixed price of a 16Gb DDR5 memory chip at $50.0 in October, up from $40.0 in June. Computer exports, driven by AI server and server-SSD demand, rose 496.6% to $5.85 billion.

Exports for January to October reached $930 billion, above $900 billion for the first time, and the ministry expects annual exports to pass $1 trillion in November. Adjusted for working days, average daily exports rose 76% to $5.24 billion.