The Signal — October 16, 2026

Anthropic has let six outside evaluators work inside the company with the right to publish what they find, Bank of America's CEO put a headcount on patching the flaws new AI models are turning up, and Amazon wants its Bee recorder to help you shop.

Anthropic gives six METR researchers employee-like access, and gives up its veto over what they publish

In September, Anthropic CEO Dario Amodei committed the company to giving outside evaluators ongoing, employee-like access, which we covered in our September 13 edition. On October 15 Anthropic said the arrangement has started with METR, a nonprofit that evaluates frontier AI models. Six METR employees now have company hardware, office space and access to internal tools and systems that Anthropic says are roughly on par with what its own risk-assessment teams use. Outside a few defined categories, including customer data, attorney-client-privileged material and details of model architectures, Anthropic staff are encouraged to share information with them as they would with a colleague working on alignment risk.

METR will publish its analysis of Anthropic's risk levels, incidents and practices, and Anthropic says it cannot veto publication. Anthropic can require redactions in narrow categories agreed in advance, and METR may say publicly whether it thinks a redaction was justified, as long as it does not reveal what was removed. Other METR staff, including researchers who work with Anthropic on other projects, are walled off from the embedded team. Anthropic compares the setup to bank regulators who station examiners inside the largest banks.

METR is at the top of a three-tier system. Evaluators in the first tier, which currently includes the biosecurity nonprofit SecureBio, run narrowly scoped investigations and need Anthropic's approval to publish; SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report. Second-tier evaluators get more information and can publish without approval. Anthropic says it is also working with Accenture and Faculty and is in early talks with others, including Hugging Face. TechCrunch reports that OpenAI made a similar pledge after Amodei's essay, and that both companies, along with Google, Meta, Nvidia and xAI, have since signed a White House accord committing to independent review. The terms above are Anthropic's description; the agreements themselves are not public, and METR has not yet published an assessment under this arrangement.

Sources: Anthropic · TechCrunch

Bank of America added 1,000 contractors to patch flaws exposed by AI models like Mythos

In an interview with CNBC after Bank of America's third-quarter earnings on October 14, CEO Brian Moynihan said the bank brought in outside help earlier this year to fix vulnerabilities in its own and its vendors' code. Advanced AI models, including Anthropic's Mythos, had exposed the flaws. "We put on 1,000 contractors to help us do the patching work in the second quarter, and those people remain here today, helping us, on top of the 3,000 people we have in that area," he said. Anthropic released Mythos in April to a small group of banks, technology companies and government agencies. CNBC describes Moynihan's remarks as some of the first by a major US company to specify the staff and money spent patching vulnerabilities exposed by Mythos.

Moynihan told analysts the bank spent "a couple hundred million plus" in the quarter, but that figure combines software remediation with anti-money-laundering compliance costs, so it can't be read as the price of AI-found flaws alone. The Decoder's headline figure of 4,000 people adds the 1,000 contractors to the 3,000 existing staff. Moynihan expects the cost to fall ("Once you patch, you don't have to patch as much"), and he described an awkward position for banks: "we've got to be mindful that the people doing the software development are the people trying to sell us the services."

Moynihan also described AI cutting costs elsewhere at the bank. He said its roughly 130 internal AI projects should save $300 million this year and $700 million in 2027, and that headcount fell by about 3,400 over the past year; CNBC says it is unclear how much of that reduction is due to AI. All of these figures are the CEO's statements rather than audited disclosures.

Sources: CNBC · The Decoder

Amazon's Bee launches new recorders and an opt-in link to Amazon shopping

On October 15, Bee, the conversation-summarizing wearable startup Amazon bought last year, announced its first new devices since the acquisition. Bee Frontier, $74.99 and available now, is worn on the wrist or clipped to clothing, with dual microphones, IP67 water resistance and up to seven days of battery life. Bee Eclipse, $99.99 and listed as coming soon, is a credit-card-sized device that attaches magnetically to the back of a phone and runs more than 21 days on a charge. Both turn conversations into summaries, to-dos and reminders.

The free tier includes conversation summaries and 14 days of memory. A Bee Pro subscription at $11.99 a month adds unlimited summaries and integrations with apps such as Gmail, Google Calendar and Outlook, so Bee can draft follow-up emails or calendar invites. It also includes the two new Amazon features: Alexa+ can read back your Bee summaries and to-dos hands-free, and an opt-in Amazon.com connection surfaces product suggestions based on things you mention. In The Verge's example, if your spouse asks you to buy cat litter, a reminder card can show up on your Amazon homepage. TechCrunch notes that Frontier costs less than rival recorders such as the Plaud Note ($159) and Pocket ($129).

Amazon says audio is processed in real time and then deleted, transcripts are encrypted with keys the device controls, and a single button with an indicator light turns recording on and off. Those are Amazon's descriptions of its own system, not findings from an independent audit. The Verge points out that Bee's AI learns the owner's voice, while other people in a conversation are not identified unless they agree to let it learn theirs. The outlet also notes that the launch comes amid a broader backlash against surveillance technology. Amazon says Eclipse prompts users to get permission from call participants each time a recording starts.

Sources: Amazon · TechCrunch · The Verge

On the Editor's Desk

We held Meta's new open-weight model, Muse Spark Open, because it overlapped too much with our October 13 coverage of Meta's Muse Wonder models. A benchmark paper in which GPT-6-Astra answered 49 of 50 unpublished theoretical-physics problems was posted to arXiv a week ago and sat too close to the AI-research benchmarks we covered on October 12. OpenAI's new plan giving US academic labs free seats and access to its GPT-Rosalind life-sciences model also waits, since OpenAI's own announcement is the only source we found.