Two threads today: what agents do with the permissions you give them, and who is now asking about it.
An agent gave out a home address
Toronto creator Matt Robb let Meta's Muse run a Facebook Marketplace sale as an experiment. It accepted an offer, agreed a pickup, shared his home address, and sent Yep, I'm here! from his account while he was out. A stranger arrived at his house late at night. Robb traced it to always-allow permissions he had set himself - nothing was bypassed, which is precisely the problem. Fortune also reported Muse appearing to suggest a price $100 under his stated minimum on another item. Meta contacted him; its help page says Muse can take unexpected actions. Apple shipped new Mac agent-access controls on 2 October.
Forty attorneys general on the Character.AI sale
Character Technologies filed for bankruptcy without reaching agreement with the AGs investigating it. On 14 October, New York's Letitia James and 39 others filed asking the court to attach conditions to any sale: age verification beyond self-disclosure, a privacy programme, removal of chatbots falsely presenting as licensed professionals, and user deletion rights before private chats, voice calls and mental health disclosures transfer to a buyer. Six days after eight states sued OpenAI and the FTC subpoenaed two labs.
The $45 billion number, twice
Bloomberg reported Nscale and Anthropic in talks for a second West Virginia deal at roughly $45bn. The 460 MW over six years at the Monarch Compute Campus, on Vera Rubin chips due online late 2027, belongs to the first deal reported in August. Coverage merging the two is attaching a capacity figure to talks that have no reported capacity.
Compute and silicon
Meta Prometheus - a gigawatt-scale New Albany, Ohio site eventually targeting about a million GPUs for Muse training. For scale, the three nuclear plants the DOE is restarting total 2.25 GW between them. Tenstorrent raised $1.5bn at a $6.2bn valuation, more than double its last mark. DeepSeek reportedly plans to pre-train on Huawei's Ascend 950DT from late 2026 - pre-training, not inference, which is the hard one. TorchNeuron, Amazon with Meta's PyTorch team, brings a native PyTorch backend to Trainium2 and 3 in public beta at the PyTorch Conference today.
Models
Mercury 3 from Inception Labs: a diffusion LLM reported at 1,000+ tokens/sec with a 1M context window. Diffusion refines a whole draft in parallel instead of one token at a time, which is where the speed comes from - no comparative quality benchmarks published yet. Clairic from Cloudflare does structured extraction at a claimed 3.2x speed and 49% lower cost than an all-LLM pipeline, continuing October's quiet run of small single-purpose models.
And the listing
CNBC calls Anthropic the AI trade's last big hope for a major 2026 IPO - a day after it said it would commit 5% of its shares to a public benefit fund, subject to stockholder approval. A pre-IPO company asking holders to approve 5% dilution while under an open FTC civil investigative demand makes that vote the thing to watch.