DeAI Daily Brief — 8 November 2026

Today in DeAI: xAI turns on default Grok API watermarking, Anthropic open-sources a prototype for proving which model produced a reply, and Buoyant ships a Kubernetes control plane for vLLM fleets.

Grok API text watermarking is on by default

As of November 7, 2026, text output from seven current Grok model variants (grok-4.7 down through grok-4.3 and the grok-4.20 family) carries an invisible statistical watermark on every xAI API endpoint. Integrators can read a per-request x-watermark header; the opt-out belongs to team admins, applies per model, and requires confirming a legal exception. xAI's docs name EU AI Act compliance as the rationale, three weeks before the Article 50(2) marking deadline. Why it matters: it is the first confirmed default-on watermark policy at a major frontier API, and the first hard answer to how providers will meet the December 2 deadline. (xAI developer documentation) Our full coverage

Anthropic open-sources FlashProve, a model-identity proof prototype

Anthropic published FlashProve as open source with code, a technical report, and formal proofs. The prototype lets anyone check that an AI reply came from the model a provider says it used, and the company claims the check can run against a 400-billion-parameter model on a single H100. Why it matters: verifiable inference is the trust primitive for every deployment where the runner is not the model owner, and the single-H100 figure is Anthropic's claim until the technical report's benchmark setup is checked. (The Next Web)

Buoyant ships Kinference for vLLM on Kubernetes

Buoyant, the company behind Linkerd, released Kinference under Apache-2.0. The control plane watches vLLM KV-cache events and engine metrics to route each request to the best endpoint, with declarative per-tenant policies and a gRPC interface speaking OpenAI chat-completions format. The vendor's own benchmarks claim up to 95% lower median time-to-first-token versus plain Kubernetes serving, with methodology in a companion post and no independent replication yet. Why it matters: self-hosted multi-tenant inference is getting real fleet-operations tooling, which strengthens the self-hosting option on the ops axis. (GitHub)

Open-vs-closed cyber gap keeps narrowing, per Vals AI benchmark

Open-weight DeepSeek V4.1 Flash hit 72.8% of attack objectives at roughly a tenth of the cost of GPT-6 Sol Daybreak's 80.8%, according to a Vals AI cyber benchmark reported by The Decoder. The gap between open and closed models on security tasks continues to compress. Why it matters: security-capability parity at a fraction of the price changes both the buying math and the risk math for open-weight deployments. (The Decoder)

Artificial Analysis Index v5 leans on secret test data

Artificial Analysis's Intelligence Index v5 is live, and reporting points out it uses hidden test data while docking open-weight models up to 16 points relative to closed models. Why it matters: every leaderboard-driven purchase decision inherits the index's comparability choices, so the open-weight penalty is worth reading as methodology, not verdict. (Trending Topics)

Watching tomorrow

Web Summit Lisbon opens November 9, and Akash's AKT-to-Solana migration portal is cited to open the same day; both are the week's next fixed dates.