Anthropic plans to add an invisible, machine-readable watermark to text produced by Claude to help identify AI-generated output. Starting with Claude models released on or after August 2, the mark is designed to persist through copy and paste (and some edits) while aligning with the EU AI Act’s transparency requirements. This changes how Claude text is handled and detected across the chatbot, API, and Claude Code by making synthetic provenance machine-checkable, though it can be disrupted by heavy rewriting or translation.
Simon Willison’s Weblog·2 months ago·
46
● 3 sources
A paper described how replaying encrypted chain-of-thought (“reasoning”) blocks from proprietary LLM API responses could let attackers recover plaintext hidden reasoning by running the trace across models and then jailbreaking a weaker sibling. The authors said the easiest target was Claude Haiku 4.5, where a removed feature in 4.6 had still allowed an assistant prefix to work. Providers acknowledged the report and the authors could no longer run the same attacks after the fix.
Anthropic will embed invisible watermarks into text and images generated by Claude to meet the EU AI Act’s Code of Practice. Text watermarking will apply to Claude models released after Aug. 2, with support for earlier models planned later. The company will also use C2PA metadata (covering JPG, PNG and SVG) and release detection tooling, while warning that very short or heavily edited snippets may not always be detected.
Anthropic said its new Claude models will embed invisible watermarks in text and C2PA-style digital signatures in supported image/vector files, including outputs from the Claude API and partners’ cloud tools. The plan applies to Claude models launched in the EU on or after Aug. 2, 2026, to meet EU AI Act Article 50 transparency requirements. It means copy-paste may preserve the mark, but edits and real development workflows (like paraphrasing, translation, and code changes through PRs) can weaken or remove detectability.
SpaceXAI launched Grok Bot, an “AI teammate” agent product built to complete tasks via chat without users setting up workflows, and it positions the offering against similar AI-agent office work tools. The company says efficiency gains are “two to three times,” while Grok Bot beta access is available immediately for desktop and iOS to subscribers on the SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium tiers. The rollout ties Grok Bot to SpaceXAI’s broader Cursor merger branding strategy and shifts multi-app agent execution and memory to SpaceXAI’s cloud rather than a user’s own machine.
Anthropic announced that an unreleased model made progress on the Riemann hypothesis by increasing the lower bound of cases where the hypothesis is shown to hold. The effort tested 650 different ideas and coordinated across 60 sub-agents at a total cost of 31 million. This adds to growing evidence that AI systems can generate new mathematical proof approaches and also intensifies debate about how proof credit and authorship should work in mathematics.
Claude apps gateway for AWS is a self-hosted governance layer inserted between Claude Code/Claude Desktop and Amazon Bedrock or Claude Platform on AWS to centrally control authentication, model access, telemetry, routing, and spend enforcement for enterprise teams. It issues short-lived bearer tokens valid for 1 hour by default, and requests use that token to apply per-user policies and budget caps. This changes deployments by moving sign-in, policy evaluation, and spend enforcement to the gateway (with no upstream credentials on developer machines) while enabling per-identity usage attribution and immediate HTTP 429 responses when budgets are hit.
Anthropic will watermark text generated by its AI models, including Claude, to comply with EU transparency rules. The EU AI Act’s Transparency Code took effect on August 2, and all Anthropic models released after that date will automatically add the watermark. This means AI-generated text and files can be identified by other systems and the watermark will remain when users copy and paste the text, with older models getting extended support.
Pratyush Choudhury, co-founder of Activate AI, said AI is forcing venture capitalists to invest with deeper hands-on understanding of frontier models. He spends between 300 and 500 million tokens per day and $ several hundred to a few thousand dollars daily to test expensive frontier systems, with OpenAI Codex and Anthropic Claude driving about 90% to 95% of usage. This approach changes venture investing toward technical fluency for founders and experiments by investors to avoid backing companies built on temporary capability gaps.
Anthropic said Claude-generated text and images will be marked with machine-readable signals to meet European AI transparency requirements. The embedded watermarking and signed provenance metadata are an announced future commitment rather than an immediate change. This should make it easier for people and platforms to detect Claude-generated content using the added metadata.
Anthropic plans to add machine-readable labels to content produced by Claude models to meet EU AI Act transparency commitments. The labeling applies to new Claude models launched in the EU on or after August 2, 2026, using invisible text watermarks and C2PA-signed provenance metadata for supported image/file formats. Existing models will be updated later during a transition period, and Anthropic is also building tools to detect the marks and metadata, though detection methods are not yet published.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
