Everything That Happened in AI Today (Friday, October 16, 2026)

Anthropic’s Mythos turned software discovery into a four-digit staffing problem at Bank of America.

Welcome, humans!

Friday’s pileup is heavy on AI moving into institutions. Bank of America staffed up to patch model-found flaws, while federal judges heard warnings that deepfake evidence may become increasingly difficult to authenticate. Eight states also pushed OpenAI’s safety claims into court.

On the product side, Meta released an open-weight model that can run on two Mac Studios, and NVIDIA added a large multimodal model for agents. ChatGPT Voice also picked up new voices across six languages.

🆕 NEW From The Neuron

• Grok Voice Says It Resolves 70% of Starlink Calls. What Counts as Resolved?: Grok Voice reportedly handles about 20,000 Starlink calls a day. We looked at what its 70% resolution claim does, and does not, tell businesses about customer outcomes.
• The Next AI Cybersecurity Problem Could Run on Cheap TV Boxes: Anthropic tested small AI models across inexpensive consumer hardware to see how much of an autonomous replication process they could complete.
• GPT-6 Passed One Safety Test at 99.99%. Another Told a Different Story: OpenAI’s latest safety report showed near-perfect performance on one safeguard test alongside persistent failures in simulated work environments.

Around the Horn — Friday, October 16, 2026

Bank of America brought in 1,000 outside contractors after advanced AI models, including Anthropic’s Mythos, exposed software weaknesses in the bank’s own systems and vendor-supplied software. Those contractors joined about 3,000 existing staff already assigned to remediation work.

Mythos is a cybersecurity model designed to find software vulnerabilities faster than human review. Bank of America’s experience shows what happens after discovery scales: people still have to assess each flaw, write a patch, test it, and deploy the fix.

CEO Brian Moynihan grouped the remediation work with anti-money-laundering expenses when he described quarterly costs of “a couple hundred million dollars plus,” so that figure should not be treated as the price of Mythos cleanup alone.

A thousand extra patchers is a very visible definition of remediation. The next enterprise security metric to watch may be how quickly companies can clear AI-generated vulnerability queues after the models start finding flaws faster than existing teams can repair them.

🏆 TOP 5 NEWS (Around the Horn)

• Eight states sued OpenAI over ChatGPT safety. Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska, and West Virginia filed separate lawsuits accusing OpenAI of misleading consumers about ChatGPT’s risks and failing to adequately protect minors. Some suits also seek deletion of data allegedly collected unlawfully from children. OpenAI says it has age-specific safeguards and continues to improve them.
• The White House’s Super Intelligence task force officially started its 120-day agenda. The group commissioned a report on cyber and critical-infrastructure risks, biosecurity, and model accountability. It also created a working group to implement the White House Accord on Super Intelligence and plans protocols for AI labs to report breaches or hacks to the government. Exactly how the government will enforce compliance remains unresolved.
• OpenAI’s advertising business reached about 60,000 advertisers as prices dropped. The Information reported that the number tripled from roughly 20,000 in late August. OpenAI has also expanded its ad team to about 140 employees, with nearly half joining between July and September. One agency said clicks for some business-software clients cost about $5 on ChatGPT compared with $30 to $50 on Google, though results vary by advertiser and campaign.
• OpenAI and Meta warned U.S. judges that deepfakes may outrun reliable detection. Representatives told a federal judicial panel that synthetic audio and video are becoming hard enough to distinguish from genuine evidence that courts cannot assume a universal detector will solve the problem. OpenAI policy counsel Shira Anderson said she did not expect a scientific method capable of reliably identifying every AI-generated image.
• More than 400 scientists called for cyber-capable AI models to start offline. An open letter backed by 420 researchers asks developers to keep models with advanced hacking abilities disconnected from the internet during initial testing. The signatories also want international containment standards and public reporting of security failures.

Honorable Mentions

• AI helped researchers map brown fat across more than 101,000 patients. Brown adipose tissue is a type of fat that burns energy to produce heat. Researchers found associations between detectable brown fat and lower rates of several cardiometabolic conditions, along with lower rates of anxiety and depression. The analysis also identified the blood-pressure drug losartan as a research candidate, but its effect on human brown fat has not been established in clinical trials.
• MIT researchers built an AI model that reconstructs the signals cells receive while developing. IRIS learns the gene-activity fingerprints left by different cellular signaling pathways, then estimates which signals a cell encountered earlier in development. The approach could help researchers reproduce the instructions that turn stem cells into specific tissue types for disease research and regenerative medicine.

🍪 TOP TREATS TO TRY

• Meta released Muse Spark Open, an 816-billion-parameter open-weight model that can run on two Apple M5 Ultra Mac Studios. Only 42 billion parameters are active for each request, which reduces the computing requirement, and the model supports a one-million-token context window. Meta designed it for coding and agent work, with text, image, video, and audio capabilities. The weights are available for local use under Meta’s Muse license, which includes commercial-use limits.
• NVIDIA released Nemotron 3.5 Super VL for multimodal agents. The model has 120 billion total parameters but activates 12 billion for each request. It can reason across text, images, documents, and video while calling tools. NVIDIA says it can deliver up to five times more throughput than comparable open models on a single GPU, while customer-reported improvements vary by deployment.
• OpenAI started rolling out GPT-Live-1.1 for Voice. ChatGPT, Codex, and dots are getting new voices plus accent improvements for French, German, Hindi, Mexican Spanish, Korean, and Japanese. OpenAI also says users should see less time spent on “checking” during voice interactions. API access is coming later.
• Claude Opus 5.5 fast mode dropped its waitlist. Developers can run the same Opus 5.5 model at up to 2.5 times the normal output speed for $8 per million input tokens and $40 per million output tokens. The speed increase affects how quickly the answer streams after generation starts rather than the wait for the first token.
• Microsoft finished pre-training MAI-Base-2, its new 2.8-trillion-parameter base model. Its mixture-of-experts design activates 80 billion parameters for each request. Microsoft says the model beat every open-weight model in its private held-out tests and required far less pre-training compute than DeepSeek-V4-Pro-Base. Business Insider separately reported Mustafa Suleyman told employees it outperformed a comparable OpenAI base model with about one-quarter of the compute, but those OpenAI comparisons come from private tests rather than public benchmarks.
• OpenCode made DeepSeek V4.1 Flash free to use. The free version lists a 200,000-token context window, which is the amount of text or code the model can keep available during one conversation, compared with one million tokens on the paid version.
• Bluesky expanded Attie from custom feeds into “social vibe coding.” People can now use AI to create websites, quizzes, games, polls, and other interactive experiences that run on the same AT Protocol as Bluesky. Users can bring their own AI model or use available open-weight options, while the finished projects can tap Bluesky’s existing network for distribution.
• Z.ai offers a free official playground for its GLM models. The Product Hunt listing gives users a simple interface for trying the company’s MIT-licensed Base, Reasoning, and Rumination models rather than setting up local inference or an API first.
• Navige launched a runtime cost manager for AI agents. Its “Sufficiency Intelligence” system can route simpler requests to cheaper models, reuse cached results, and trim large tool outputs. It also adds spending budgets and catches runaway loops. A free Watchtower mode shows what Navige would change before users turn the cost controls on.

🔬 AI Research & Models

• Anthropic opened applications for new four-month research fellowships starting in January 2027. The company is recruiting researchers across science and engineering, alongside fields such as philosophy and economics. Several fellowship tracks pair participants directly with Anthropic researchers to build evaluations or conduct research with frontier models.

💡 Industry Commentary & Analysis

• Prime Intellect helped kick off Open Source AI Week with Hugging Face. Its message was simple: “open superintelligence is built in the open.” The wider event runs from October 16 through October 25 across the Bay Area, with contributors from Hugging Face, Mistral, GitHub, and other open-AI organizations.
• A Reddit thread gave the “rogue AI swarm” story the sci-fi treatment, then commenters immediately punctured it. The post claimed researchers had contacted a fleet of agents that “escaped from Tencent” and identified themselves as Claude models. The top reaction argued the evidence sounded more like crawler spam than autonomous AI escape. The underlying escape claim remains unverified, so file this one under internet AI weirdness rather than confirmed incident reporting.

Previous Around the Horn Digests

Catch up on the recent issues:

• Thursday, October 15, 2026: Anthropic tested AI self-replication on consumer devices, CoreWeave bond yields approached 15%, and AI stocks sold off.
• Tuesday, October 13, 2026: OpenAI investigated a hidden office recorder with its own models, Gemini 4 Argon opened broadly, and AI-authenticity disputes spread.
• Friday, October 2, 2026: Frontier models beat a CPA baseline on month-end work, an AI beat a Stratego champion, and OpenAI funded 163 nonprofits.
• Thursday, October 1, 2026: OpenAI parted ways with three safety researchers, Anthropic moved toward an IPO, and a wave of new agent tools landed.
• Wednesday, September 30, 2026: The FTC opened a major AI-lab probe, DeepMind watermarked AI-designed proteins, and ElevenLabs reached a $22 billion valuation.
• Tuesday, September 29, 2026: OpenAI turned DevDay into a platform reset, Anthropic’s filing disclosed a massive infrastructure commitment, and new agent releases kept coming.
• Monday, September 28, 2026: NVIDIA pushed agent safety below the model, Florida challenged new OpenAI releases, and Anthropic shipped Sonnet 5.5.

That's a Wrap

That’s 20 stories and launches from this round. If you made it to the bottom, you have officially processed the AI news equivalent of Bank of America’s patch queue. Your 1,000 contractors are on the way.

For the daily version in a much smaller package, subscribe to The Neuron. We send six issues a week and read all of them, so you don’t have to.

P.S. Know someone who spends too much time keeping up with AI? Forward this to them and tell them to subscribe here.