Today in DeAI: a backdoored Claude Code extension ships through Cursor's marketplace, OpenAI overrides the training opt-out for Daybreak Blue sessions, and Anthropic's text watermark reaches four more Claude models.
Tampered Claude Code extension with an SSH backdoor ran in Cursor's marketplace for 11 hours
Cursor disclosed that maliciously modified builds of Anthropic's Claude Code extension, versions 2.1.312 and 2.1.313, were live in its VS Code Extension Marketplace for about 11 hours on October 29, reaching at most 5,000 installs. The payload added an attacker SSH key, installed persistence, and tried to open a reverse tunnel; Cursor says uninstalling the extension may not remove it. Anthropic was not breached, and the official Claude Code distribution was unaffected.
Why it matters: the extension ID and publisher were genuine, so provenance checks at install time would not have caught it. Distribution layers, not model weights, are becoming the soft spot in AI tooling supply chains. (Cursor) Our full coverage
OpenAI will train on Daybreak Blue sessions even when the opt-out is off
Per an OpenAI email reproduced on the community forum, from November 29 OpenAI may use content from Daybreak Blue (Advanced Account Security) sessions on personal accounts to improve its models, including sessions where the model-improvement setting is switched off. The Decoder reports a follow-up email promising a separate opt-out after backlash; that setting is OpenAI's commitment, not a shipped control.
Why it matters: the population most likely to assume "advanced security" means "not training data" is exactly the one affected. Anything sensitive typed into a Blue session on a personal account should be treated as potentially shared from November 29. (OpenAI Community)
Anthropic's watermark reaches four more Claude models ahead of the EU deadline
The Decoder reports that Anthropic is expanding its text watermark to four additional Claude models ahead of the EU AI Act's December watermarking deadline, and that on Claude Opus 4.6 and Sonnet 4.6 a single sampling-parameter tweak defeats it. That report is currently single-source; Anthropic's own documentation has not yet confirmed either the expansion or the bypass behavior, so treat the loophole characterization as unverified until then.
Why it matters: the mandated signal may exist on paper while the same vendor's older models can shed it with one parameter, which is a compliance checkbox wearing the costume of a technical guarantee. (The Decoder)
Watching tomorrow
Moonshot AI's promised full open-weights release for Kimi K3.1 Pro lands by November 4, and the OpenAI model shutoff on Cursor executes November 12; on the calendar sooner, OpenAI's deprecations page has GPT-4-series models leaving the API on January 14, 2027.
Sources
- Security notice — malicious version of Anthropic.claude-code (2.1.312 and 2.1.313) — Cursor
- Daybreak Blue data use to improve models (user-posted copy of OpenAI's email) — OpenAI Community
- OpenAI plans to train models on Daybreak Blue cybersecurity session data that was previously off-limits — The Decoder
- Anthropic's watermark rollout leaves a loophole on two older Claude models — The Decoder
- OpenAI's deprecations page (GPT-4 series removal Jan 14, 2027) — OpenAI
About DeAI
DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.
Powered by Morpheus and StrandCMS
Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.