Today's Headlines

  • Anthropic says it will embed an imperceptible watermark in the text its models write — no detection mechanism has been published yet
  • Encrypted reasoning traces can be recovered in plaintext — 704 secrets from 6,708 public logs, 64 of them absent from the visible session
  • Spotify will badge AI Personas and drop them from recommendations — self-disclosure opened August 11, badges appear in mid-September
  • OpenAI's unreleased Astra model produced solutions to ten open mathematics problems — the announcement was rewritten over questions of credit
  • Amazon is funding a large gas plant in Texas — permitted for up to 33 million tons of carbon dioxide a year

What the three stories today have in common is the invisible side. Not the content itself, but how its origin gets shown — that is the side all three are working on.

What Anthropic is moving is where the mark goes. It says the watermark sits in the text itself, so that it travels along when the text is copied and pasted somewhere else.

What the researchers have shown is what comes back out of the side that was meant to stay hidden. The reasoning blocks that frontier APIs return in encrypted form can be turned back into plaintext.

What Spotify is moving is a mark meant for human eyes. Artist identities built by AI will carry a badge that says so.

Today's Top Three

Anthropic says it will embed invisible watermarks in Claude's text

Anthropic has announced that it will embed an imperceptible, machine-readable mark in the text generated by Claude and its other models.

The mark goes into the text itself. By the company's account, the watermark is embedded directly into the text, so it travels with the text when it is copied and pasted elsewhere, and may persist through some editing.

Files are handled by a different mechanism. For supported file types — .svg, .png and .jpg — the company says it attaches signed provenance metadata, following the C2PA (Coalition for Content Provenance and Authenticity) open standard also used by Adobe, OpenAI and Google.

When it applies depends on the generation of the model. The company states that "Claude models launched on or after August 2, 2026 will support machine-readable marking at launch," and says support for models released before that date is still in progress.

The scope runs across products. Marking applies to Claude Platform (API), Claude, Claude Code, Claude Cowork and Claude Tag, including usage through AWS, Google Cloud and Microsoft Foundry, wherever Claude is offered.

This is the point most easily misread today. The company says it will "share details on detection mechanisms in forthcoming technical documentation," which means there is currently no published way for anyone else to test a piece of text for the mark. This is an announcement that marking is being introduced, not a tool that anyone can use.

Even once detection arrives, what it can tell you has a ceiling. The company says a detected mark indicates that content may have been processed by Claude and is not fully conclusive, because people also use Claude to edit, translate and summarise material they did not write with it.

The reported context is regulatory. TechCrunch and The Verge both frame the move as compliance with the transparency requirements of the EU AI Act, which took effect on August 2, and The Verge notes that the law gives existing products a four-month grace period.

Encrypted reasoning traces recovered in plaintext — 704 secrets from public logs

Researchers have published a method for turning the encrypted reasoning that major LLM APIs return back into plaintext, and used it to recover 704 secrets from publicly available agent logs.

The method starts from the portability of those blocks. According to the research team, the APIs from Anthropic, OpenAI and Google return a model's reasoning to the client as an encrypted block, which the client sends back when the conversation continues. Those blocks can be replayed outside their original context, across sessions, users and models.

Reading them out takes two API calls. A block produced by a frontier model is fed to a weaker sibling model from the same provider, that weaker model is jailbroken, and it is asked to transcribe the attached reasoning verbatim.

The strong model is never touched. The team states that this recovers the stronger model's hidden reasoning in plaintext without attacking that model directly and without triggering its anti-distillation safeguards.

The material came from the open web. The team collected 6,708 publicly available agent trajectories from GitHub and Hugging Face that still contained encrypted reasoning blocks, applied its decoding pipeline to every signed block, and reconstructed 315,320 reasoning blocks.

What came out of them is concrete. Restricting the count to genuine, non-benchmark user sessions, the team recovered 704 distinct privacy artifacts, including 62 API keys, 33 passwords, 24 access tokens and 30 personal email addresses, alongside names, postal addresses and internal URLs.

One figure carries most of the weight for companies. Of those 704 artifacts, 64 appeared nowhere in the visible session and existed only inside the hidden reasoning blocks. Anyone who shared such a log and reviewed it by eye had no way of finding those 64.

The paper is public. "Stealing Reasoning Traces from Proprietary LLM APIs" was posted to arXiv on August 10, by researchers at the ELLIS Institute Tübingen, the Max Planck Institute for Intelligent Systems, the Tübingen AI Center, MATS Research, the University of Tübingen and Snyk.

Spotify will badge AI Personas and exclude them from recommendations

Spotify has announced that artist identities generated by AI will carry an AI Persona badge and will be excluded by default from its recommendations.

What the badge conveys is about identity. In the company's words, the badge signals to listeners that an artist's identity may be AI-generated and does not represent a real person.

The rollout comes in two stages. Artists can self-disclose through Spotify for Artists starting August 11, and the badges become visible to listeners starting in mid-September.

The recommendation change bites harder than the label. The company says that by default it will not include AI Personas in any editorial or algorithmic recommendations, so a flagged persona reaches only the listeners who follow it directly.

The system does not wait for artists to come forward. Spotify says it will review artist profiles and identify those whose public identity appears to represent photorealistic AI-generated identities, beginning with profiles that meet defined audience thresholds.

There is a process for the artists on the receiving end. When Spotify applies a badge, the company says the artist will be notified and given the opportunity to self-disclose or appeal.

The boundary is stated plainly. The company says the badge "is about the artist's public identity, not about how the music was made," and points to separate tools — AI Credits and SongDNA — for transparency about the creative process itself.

Other Developments

Models

A follow-up. The open-weight release of Qwen3.8-Max was expected today, August 12, and as of today the weights are not on the official Qwen account on Hugging Face. The most recently updated model on that account dates from July 22.

A follow-up. On Meta's open-weight turn, covered in yesterday's edition, Ars Technica reports that the company has promised to open the weights of its more powerful Muse Spark 1.2 model in the next few weeks. Muse Spark 1.2 itself was released on August 5, alongside Muse Code, a terminal coding agent.

Products

A developer has published an account of intercepting GitHub Copilot's traffic in VS Code with mitmproxy. In the test, a fake credential placed in a .env file was sent along with a completion request made while editing an unrelated pyproject.toml, a file the developer had never touched with the secret in it. The author explains that the default behaviour of sending recently edited files as context is hardcoded at up to 20 files, 8 edit summaries and 3 lines around each change, so switching completions off for particular extensions does not prevent it. The account also shows that conversation history is stored in plaintext in a local SQLite database, session-store.db, with no redaction, sanitisation or secret filtering anywhere in the write path.

Knowledge Sense, a Tokyo company, has added a feature to the Notebook function of its enterprise AI service ChatSense that pulls PDF and Office files directly from Box as sources, without saving them to a local machine first, and carries them straight through to slide generation. The company says corporate users who manage documents in Box had asked for it. ChatSense is used by more than 500 companies, including firms listed on the Tokyo Stock Exchange Prime market.

Research

The Verge has reported on how mathematicians are taking OpenAI's announcement that its unreleased Astra model produced solutions to ten long-standing open problems. The problems span sphere packing in higher dimensions, error-correcting codes, the limits of how complex connected networks can get before structural patterns emerge, quantum game theory and the existence of non-sofic groups. On that last result, Francesco Fournier-Facio of the University of Cambridge and others said the announcement minimised the contributions of Andreas Thom and Gábor Kun, whose recent work laid the groundwork. OpenAI has since changed its wording from sharing "results to problems that have been open and have seen no progress on the main result for at least a decade" to "results, each of which resolves or makes substantial progress on a long-standing open problem," and a company spokesperson told The Verge the language was updated to better reflect the prior research the results build upon. OpenAI released more than 250 pages of papers, a further 60 pages on how the ideas came together, and certified each result with the proof assistant Lean.

Ars Technica has reported on academic peer review buckling under a surge of submissions and a shrinking pool of reviewers. The number of papers indexed in Scopus and Web of Science is growing at 5.6 percent a year, and by one estimate researchers worldwide devote a collective 15,000 years of work to peer review annually, worth $1.5 billion for the US share alone if it were paid. An editor at Human Immunology says he recently had to email around thirty researchers to find a single reviewer, where five years ago five to ten emails would have secured three.

Researchers affiliated with the University of Chicago have published a study in PNAS based on experiments with roughly 2,000 parents of teenagers. What parents said they would pay for a child's AI subscription rose by more than 60 percent — from about $17 to about $28 — simply when the stated share of local teenagers using AI moved from 20 percent to 80 percent. Telling parents about the downsides barely moved the amount they would pay, while the share saying they would prefer a world in which no student could use AI rose from 44 percent to 57 percent. Among those who favoured a full ban yet still let their own child use it, the most common reason, given by about one in five, was not wanting to be left behind.

Business

OpenAI has completed a $7 billion tender offer letting employees cash out shares, TechCrunch reported, citing Bloomberg. The valuation was $852 billion, matching its March 2026 funding round. The company filed confidentially with the Securities and Exchange Commission in June in preparation for a possible IPO, and had not responded to a request for comment by the time TechCrunch published.

Amazon has confirmed that it is investing in a natural gas power plant to supply a large planned data center in Pecos County, Texas. The New York Times reported that the plant could become the largest single source of climate pollution in the United States, and under the current plan it would be permitted to release 33 million tons of carbon dioxide a year. Amazon told Ars Technica that it remains committed to its 2040 net-zero Climate Pledge, and described the project as on-site generation that will not raise electricity costs for Texas families.

Discovered Materials, a startup searching for new materials to cool chips more efficiently, has raised a $9 million seed round led by Lightspeed India Partners, with participation from Peak XV Partners and angel investors including Paul Graham. The company has built a pipeline in which swarms of AI agents generate candidate materials and physics models it trained itself verify them in simulation. One co-founder said that where his doctoral work produced around 20 guesses a day, running agents around the clock now produces thousands.

Source: selected by the editorial desk from the AI news inbox (22 items collected on August 12, 2026; 1 primary, 21 secondary).