Tuesday, October 13, 2026

TLDR

Meta releases Muse Wonder, its first frontier model, with API access at $4/$20 per million tokensA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ (Oct 12) Pricing
Character.AI files for Chapter 11; Disney buys some assets for $20M (Oct 13) Business
GitLab patches CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 10.0 unauthenticated RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ (CVE-2026-97229) in the CI/CD regex path (Oct 13) Security
Amex proposes an "Intent Intelligence" framework to stop AI agents making unauthorized purchases (Oct 13) Agents
New paper shows hidden model capabilities can resist elicitation-based audits (Oct 9, trending Oct 13) Policy Trending

Things your engineering team should look at or do this week.

If you self-host GitLab, upgrade to 19.2.8, 19.3.5 or 19.4.2. CVE-2026-97229 is unauthenticated and CVSS 10.0. No exploitation has been reported yet, but treat it as urgent.
Run Muse Wonder against your own eval set before considering it. Meta's comparisons to Fable 5.1 and GPT-6 Astra come from its own model card. Pricing is on par with GPT-6 Sol and Opus 5.5.
If you build agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ checkout or payment flows, read Amex's intent-capture design. A signed "proof of intent" token is a pattern worth borrowing for any agent action with financial side effects.

Meta Muse Wonder / Wonder Full (Oct 12) Pricing — Meta Superintelligence Lab's first frontier model, stronger than Muse Spark at coding and complex tasks, and it emphasizes "active memory." Muse Wonder is on the Meta Model API at $4 in / $20 out per million tokens; Wonder Full is limited to Muse Code subscribers and not on the API. Open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ status wasn't stated. Benchmark claims are Meta's own. (TechCrunch)
Janas: run MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ LLMsLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ larger than RAM by streaming experts from SSD (trending Oct 13) Open Source Infra — A dependency-free C engine that keeps experts on NVMe and caches hot ones. The author reports about 20–25 tok/s for Qwen3-Next-80B on a 32 GB laptop. It is an experiment tested on one machine, Linux x86-64 only, GPL-3.0+. (GitHub)
Amex Intent Intelligence Framework (Oct 13) Agents Customer AI — A Create Intent API records the user's constraints, the user authenticates, and the agent gets a Proof of Intent Token and a derived payment credential, with a log of what was authorized. Amex is talking to payment partners about making it a standard. (TechCrunch)

GitLab CVE-2026-97229, CVSS 10.0 (Oct 13) — An integer overflow lets an unauthenticated attacker run code on the server through a crafted regex in CI/CD configuration. Fixed in 19.2.8, 19.3.5 and 19.4.2; no workaround was described. (The Hacker News)
Hidden capabilities that evade audits (Oct 9) Policy Trending — Researchers embed encrypted sub-networks in model weights that run only on a secret trigger. In a small proof of conceptProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗, the hidden behavior survived fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗, weight noising and gradient-based trigger search. It argues that capability evals alone can't certify an untrusted model's weights. (arXiv)

Character.AI files for Chapter 11 (Oct 13) — Assets and liabilities are put at $50–100M. Disney is buying some assets for $20M plus a non-exclusive perpetual license to the technology. The platform keeps running during a court-supervised sale expected to close in late November or early December, and it will be wound down if no buyer emerges. Lawsuits and the cost of monetizing consumer chat are cited. (TechCrunch)

Amazon drops NDAs in data center talks with local governments (Oct 11) Infra — Following Microsoft's similar move earlier this year, as data center backlash and permit moratoriums (including a one-year New York moratorium) grow. If you rely on new regional capacity, permitting delays are a risk. (TechCrunch)

DishwasherBench v0.1.0 (trending Oct 13) Open Source — A physics-sim game and benchmark where LLMs load a dishwasher, playable via an MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server. The authors call early results (GPT-6.1 Sol and GPT-6 Astra scored similarly) "an anecdote, not a benchmark result." Fun for agent-harness experiments, not for model selection. (GitHub)