Monday starts with two stories that should send you back into your own code. One is a compliance detail Anthropic put in an email, the other is a five-step exploit that only worked because three harmless features met each other. After that, two releases you can run locally today: DeepSeek's agent that sits in your repo until someone asks it for work, and a decision model small enough for a laptop.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Claude's new text watermark doesn't apply on Opus 4.6 and Sonnet 4.6 if you change temperature, top_p or top_k
Anthropic told API customers it's rolling an invisible word-choice watermark, a SynthID variant, onto older models ahead of the EU AI Act's December deadline. Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 are getting it now; models released after August 2 shipped with it. The catch, per The Decoder: on Opus 4.6 and Sonnet 4.6 it only applies when all three sampling parameters sit at their defaults. You can't switch it off, and detection is a private-preview API for now.
Link: the-decoder.com
Try it now
Find out which of your calls fall outside the watermark:
grep -rnE "temperature|top_p|top_k" --include=*.{py,ts,js} .
If a hit sits next to claude-opus-4-6 or claude-sonnet-4-6 and you rely on the watermark for compliance, drop the override and let the defaults stand. If you want unwatermarked output from those two models, keep in mind that's exactly what auditors will ask about.
Story II
A ChatGPT share link could read a signed-in victim's private chats, and OpenAI paid $5,000 to close it
Parameter Labs built a chat with a graph and an HTML Preview, then shared it. The Preview sent postMessage calls to MathLive's virtual keyboard in the parent page, which never checked the sender and wrote the payload with innerHTML. That spawned an unsandboxed frame. The CSP allowed YouTube, and YouTube's oEmbed endpoint reflected script, so the code ran as chatgpt.com, listed recent chats and exfiltrated one. Reported October 18, fixed October 26.
Link: parameter.ai/labs
Try it now
Audit your own message listeners:
grep -rn "addEventListener(['\"]message" src/
For every hit, check two things before any state change: event.origin matches an allowlist, and event.source is the frame you expect. Then search for innerHTML fed by message data. Parameter Labs' own advice: never accept raw HTML as config across frames, and treat CSP as a backup, not the wall.
Story III
DeepSeek open-sources dsh-superagent, a GitHub coding agent that runs one isolated session per task
The repo, MIT-licensed and already past 11.4k stars, packages a router-worker agent for DeepSeek Harness. A long-running Router watches GitHub and turns an allowed person's request into a Ticket, and each Ticket gets its own Worker session, directory and usually a pull request. In DeepSeek's simulated test of about 300 messages, p95 completion was 15.7 minutes against 29.5 for a single session with subagents. That test was synthetic and didn't measure code quality. DeepSeek calls it a research prototype that runs with full OS-user access.
Link: github.com/deepseek-ai
Try it now
The README hands setup to your coding agent. Create a throwaway OS user, clone the repo, open your agent there and paste:
"Read SETUP.md in this repository and set up dsh-superagent with me."
Keep the accept mode on observe, the default, for a week. It only records what it would have picked up, so you can read those decisions before letting it start work on its own.
Story IV
Nace AI's Drex Max v1, a 27B decision model, runs in 4-bit on a 24 GB Apple Silicon Mac
Drex Max doesn't write text. You send a state plus typed questions (choice, yes/no or a score) and get a probability for every option in one forward pass, which fits routing, ticket triage and guardrail checks. Nace AI reports 67.52 on the public Decision Index 0.4, up from 61.25 for Drex v2.0. The affine Q4 build is 14.45 GB and, in Nace's own tests, flipped no answers against the 8-bit version. Local Mac and CPU paths are text only. The license is Nace.AI Open RAIL-M, not MIT.
Link: huggingface.co/nace-ai
Try it now
On a 24 GB or larger Mac, straight from the model card:
brew install nace-ai/tap/kapitsa
hf download nace-ai/drex-max-v1-affine-q4 --local-dir ./drex-max-v1-affine-q4
kapitsa-systemone --model ./drex-max-v1-affine-q4 --port 8000 --context 2048
Then POST the repo's examples/request.json to http://127.0.0.1:8000/v1/systemone and swap in one of your own routing questions.
— The Vibe Gate news desk. We read the firehose so you can keep building.