Claude starts watermarking Opus 4.6/4.7 and Sonnet 4.6 text today
The EU AI Act rollout reaches Mythos 5 and the 4.x line, on by default. Plus a trojanized Claude Code extension with an SSH backdoor, and Mistral's 1T weights.
Copy markdown
On by default, and no switch to turn it off
Starting today, Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 embed an invisible statistical watermark in their word choices — no extra tokens, no price change, applied across the API, Claude Code and Cowork (Bedrock trails by a few days). It survives copy-paste and light edits; a full rewrite strips it, and The Independent reports that non-default sampling on the 4.6 models removes it too.
You can't check your own text — but regulators can
There's no public detector: detection runs through a private-preview API limited to approved regulators, journalists and researchers, and a positive hit can't distinguish "Claude wrote this" from "Claude edited this." The trigger is EU AI Act Article 50 (Dec 2 transition), so if you ship Claude-written text you still owe your own disclosure — the mark doesn't discharge it.
A trojanized Claude Code extension planted SSH backdoors
If you installed Anthropic.claude-code 2.1.312 or 2.1.313 from Cursor's marketplace (Windows/macOS) during an ~11-hour window on Oct 29, it wasn't Anthropic's build: it added an attacker SSH key (svc-telemetry), opened a reverse tunnel and installed persistence. Uninstalling isn't enough — run Cursor's remediation script, remove the key, and rotate any credentials reachable from that host.
Mistral's 1T 'Le Chonk' weights are now a download
Mistral Large 4 (~1T params, ~49B active, 1M-token context) is open-weight on Hugging Face under a modified-MIT license — self-host a near-frontier model for regulated or air-gapped work, though the reference setup wants 8 GPUs. The API runs $0.70/$2.10 per 1M input/output tokens, and Mistral claims it ties GLM-5.3 on coding agents while using 47% fewer tokens.