AI News Briefing — Cisco Talos infostealer asks an LLM to rank victims

Model releases

[2026-10-31] Anthropic’s marking page now lists its pre-August models, including Opus 4.5, 4.6 and 4.7, Sonnet 4.5 and 4.6, Haiku 4.5 and Mythos 5, as carrying the invisible text watermark on Anthropic’s own apps and API. Served through AWS, Google Cloud or Microsoft Foundry, the same models are still unmarked, and none of them adds Content Credentials to files. Newer models such as Opus 5.5 carry all three marks. (official)

Which route serves the model now decides whether its text is marked, so the same Sonnet 4.6 call can come back watermarked from Anthropic’s API and unmarked from Bedrock.

For Solution Architects: A team whose disclosure policy relies on provenance marks gets them on these older models only by calling Anthropic directly; moving the traffic to a cloud marketplace quietly drops the watermark.

[2026-11-01] nv-uefi-unlocker, an open-source tool that boots from a USB stick, lifts the half-rate cap on FP16, BF16 and TF32 tensor math with FP32 accumulate on GeForce RTX 20, 30 and 40 cards. The project reports up to 2.6× on those paths, but only 19% on LLM prefill and 5% on decode, which stays memory-bound. It needs Secure Boot off or a self-signed loader, and reruns on every cold boot. (source, official)

Anyone serving a model on a gaming card should plan around the 5% decode figure, not the 2.6× headline.

For ML / Data Engineers: The big gain sits in the FP16 and BF16 tensor paths with FP32 accumulate, so a mixed-precision fine-tune on an RTX 40 card is the workload worth timing before and after, remembering the tool must run again after every cold boot.

[2026-11-01] OpenAI told Pro subscribers by email that ChatGPT and Codex will draw from one allowance, according to BleepingComputer. Unlimited Instant doesn’t count against it, while heavier models and longer tasks use more. Remaining usage and reset times sit under Settings > Usage, and running out means waiting, upgrading or buying credits. The email gave no start date and no numbers. (source)

Codex sessions and everyday chat now compete for one pool. Until OpenAI publishes numbers, nobody can say whether that leaves more room than two separate limits did.

[2026-10-31] Cisco Talos documents VOLTA, a Windows infostealer that sends each victim’s stolen data to an LLM through OpenRouter and asks it to rank the victims by what their accounts are worth. It defaults to Claude Opus 4.6; when a reply reads like a refusal, it retries on Kimi K2, then Gemini 2.5 Flash. The stage is optional and skipped without an API key. Talos first saw the sample on October 11 and suggests alerting on headless Chrome launched by processes that aren’t developer tools. (source)

Each refusal cost the malware only a retry, since one OpenRouter key reaches all three vendors’ models. An office laptop calling OpenRouter is itself worth a look.

Kimi K3.1 Pro weights: not yet on Hugging Face; Moonshot’s date is November 4.
Gemini 4 Carbon general access: still cyber defenders and partners only; no date.
ChatGPT combined Chat and Work tab: OpenAI says a small test starts in the coming weeks, and the shared ChatGPT–Codex allowance has no start date either.

A start date plus the allowance numbers would show whether Pro users gain room or lose it.
Cursor marketplace follow-up: Cursor still hasn’t said how the tampered build was published.