Today's 12 Stories — Tuesday, November 10, 2026

Funding & Investment Story 1 of 12

Anthropic's Nearly $70 Billion IPO Plan Circulates to Banks, Then Gets Pulled Back

Reuters reported on November 9 that details of Anthropic's initial public offering plan, including a plan to raise nearly $70 billion, were sent to banks and then retracted late Monday. According to emails Reuters saw, an Anthropic representative first sent a securities filing and then followed with a second message saying the pricing details were inaccurate and that the filing would be updated later.

The numbers in that first version are still striking. According to the original filing Reuters saw, Anthropic planned to offer about 867 million shares at $70 to $80 per share, which would raise $69.3 billion at the top of the range. Reuters reported that the same filing named Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC as agreeing to take part in a private placement, with $26 billion in investments lined up alongside the offering.

Reuters also reported figures from an amended filing. It showed Anthropic revenue of almost $17 billion for the three months ended September 30, 2026, and Reuters reported that the company has nearly $900 billion in computing commitments with partners including Amazon and Google.

For executives, the retraction is the detail that matters most. The pricing terms that circulated were, by the company's own account to banks, not accurate, so none of the headline figures should be treated as final until a corrected filing appears. What can already be said is that the filing's investor list reads like a map of the AI supply chain, with a cloud provider, chip designers, a memory maker and a foundry all lined up as shareholders in a model developer.

That overlap is worth watching if you buy from any of them. When your cloud vendor, your chip vendor and your model vendor share a cap table, pricing, capacity allocation and roadmap decisions can start to move together. Procurement teams that treat those as independent negotiations may want to revisit the assumption once the final filing is public.

AI Models Story 2 of 12

Anthropic Releases Fable 5.2 and Mythos 5.2, Claiming a Machine Checked Proof of Fermat's Last Theorem

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Fable 5.2 is priced at $8 per million input tokens and $40 per million output tokens, according to the company.

The most eye catching claim is mathematical. Anthropic says a team of 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in 30 hours. The company says Claude Opus 5.5, given the same setup, completed only 26% of that formalization in 18 hours. These are Anthropic's own results, from its own announcement, and have not been independently reproduced.

On software engineering, Anthropic reports that Fable 5.2 scores 69.9% on FrontierSWE v2, versus 62.3% for Opus 5.5. On security, the company says Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 in an internal CVE Rediscovery evaluation.

Read the benchmarks as a vendor's claims about its own models, chosen and run by the vendor. The direction of travel is clear, though. The gap over the previous generation is large enough that teams running agent workflows on Opus 5.5 will want to rerun their own evaluations rather than rely on public leaderboards.

The Fermat result deserves a measured reading. Formal proof checking is attractive as a test of reasoning because the output is verified by a machine, not judged by a person. Still, a 30 hour run by 1,000 agents is a compute budget few organizations will match, so the practical lesson for most executives is about what agent teams can now coordinate, not about mathematics.

The security claim cuts two ways. A model that finds more known vulnerabilities helps defenders, but the same capability is available to attackers, and the weeks ahead will show whether safeguards hold. Security leaders should treat the release as a prompt to review how their own teams use AI in code review and patching.

AI Models Story 3 of 12

Meta's Muse Ember Undercuts Rivals on Price at $0.10 per Million Input Tokens

Meta released Muse Ember on November 9, 2026, priced at $0.10 per million input tokens and $0.50 per million output tokens. Meta says the model has a 1 million token context window and up to 512,000 output tokens.

On the Artificial Analysis Intelligence Index, Meta reports that Muse Ember scores 36 at max effort, tied with Claude Haiku 5.5, versus 28 for GPT-6 Luna. Meta says Muse Ember's cost per Intelligence Index task is $0.14, versus $2.39 for its own earlier Muse Spark 1.3.

Meta highlights agentic work in particular. It reports that Muse Ember scores 37.8% on AutomationBench-AA, versus 13.9% for Claude Haiku 5.5 and 20.2% for GPT-6 Luna. Meta's own release is also candid about where it trails: it shows Claude Haiku 5.5 ahead on FrontierCode 2.0-AA, 34.9% to 27.2%.

That mix of results is the useful part. A model that wins on automation tasks and loses on coding is a candidate for a specific job, not a general replacement, and the price gap suggests where it might fit. High volume workflows such as ticket triage, document processing and routing are the natural place to test a cheap model with a very long context window.

As always, these are figures from the vendor's own announcement. The sensible next step for a technology team is a bake off on its own data, scored on its own accuracy bar, with cost per completed task as the headline metric rather than cost per token.

The broader signal is competitive pressure on pricing. When a frontier lab prices output at $0.50 per million tokens, contracts signed a year ago look expensive, and renewal conversations with other providers get easier to open.

Industry Dynamics Story 4 of 12

Nvidia's Board Opens an Independent Investigation After Reports About CEO Jensen Huang

On November 8, 2026, Nvidia's board announced an independent investigation, overseen by a committee of independent directors and assisted by outside independent counsel. The board said the step followed reporting that founder and CEO Jensen Huang was the target of a foreign intelligence effort.

In its statement the board said it has no reason to believe that either national security or Nvidia secrets were compromised. It also said Huang supports the review and has committed to cooperating with it.

The underlying reporting came from The New York Times on October 29, in an account relayed by The Hill. The Times reported that US spy agencies discovered that Alice Huang, an executive at Singapore based data center company Megaspeed, funneled information to Beijing while developing a close relationship with Jensen Huang. The Hill reported that an Nvidia spokesperson said the diversion allegations were thoroughly investigated and debunked last year.

Both sides of that exchange matter. The allegations are reporting by a newspaper citing anonymous sources, and Nvidia disputes them. The board's decision to open a formal inquiry, led by independent directors rather than management, is a governance step that shows the company is treating the matter as more than a public relations problem.

For executives, this is a reminder that concentration risk in AI includes people as well as chips. A single supplier that sits at the center of global AI spending is also a single point of exposure to political and security scrutiny. Boards of customer companies should ask whether their AI infrastructure plans assume uninterrupted access to one vendor's products and leadership.

The timeline to watch is the investigation's own. The board said it will provide further updates as appropriate, and any findings could influence how Washington treats chip exports and how large buyers weigh supplier concentration in their next purchasing cycle.

AI Infrastructure Story 5 of 12

Google Adds TPU 8tx and TPU 8i Max to Its AI Chip Lineup

Google announced two new chips on November 9, 2026: TPU 8tx and TPU 8i Max. Google says TPU 8tx delivers up to 2x lower latency than TPU 8t and is paired with 5th Generation AMD EPYC CPU hosts. It describes the chip as built for secure and confidential inference and also suitable for small scale training.

TPU 8i Max takes a different route. Google says it combines two TPU 8i chips on one package and delivers double the compute, double the SRAM and double the HBM bandwidth of TPU 8i. In practical terms, that is a way to fit more silicon into the same rack footprint rather than waiting for a new manufacturing generation.

Google also said its TPUs have nearly 100 customers, which gives some sense of how far its custom chips have spread beyond the company's own products.

What is missing is as telling as what is included. The announcement gives relative claims, such as twice the latency improvement and double the compute, but no absolute performance figures or pricing for either system. Buyers will need those numbers before comparing the chips with Nvidia hardware or with Google's own earlier generations.

The emphasis on secure and confidential inference is the part executives in regulated industries should notice. Confidential computing keeps data protected while a model is processing it, which speaks to the main objection many banks, insurers and health systems raise about running sensitive workloads on shared cloud infrastructure.

The pairing with AMD processors is also notable. It shows the market for AI infrastructure is built from many suppliers rather than one, and that cloud providers are mixing their own accelerators with third party components. For customers, that diversity could translate into more choice on price and capacity, provided the vendors publish the specifications needed to compare them.

AI Research Story 6 of 12

Nvidia Unveils GR00T 2, a Robot Model That Handles New Tasks Without Training Demonstrations

Nvidia announced GR00T 2 on November 9, 2026 at the Conference on Robot Learning in Austin, Texas. The company says the model completed 40 of 50 evaluated tasks at least once without task specific demonstrations.

With a small amount of teaching, Nvidia says performance rises sharply. The company reports that GR00T 2 reached 71% average success across 50 tasks with one demonstration per task, increasing to 95% with five demonstrations.

Nvidia also says the model was trained in part on 100,000 hours of Nvidia open source human egocentric data, meaning video recorded from a person's point of view. That detail points to a strategy: using human video, which is abundant, to teach robots skills that are expensive to gather through robot hardware.

The numbers are Nvidia's own evaluation results, so treat them as promising rather than proven. Success on a benchmark task set is not the same as reliability on a factory floor, where a 95% success rate can still mean frequent interruptions at scale.

Even so, the trend line matters for operations leaders. The cost of deploying a robot has long been dominated by programming it for each task. A model that needs five demonstrations instead of a custom engineering project changes the economics of automation for smaller production runs, warehouse variation and tasks that change often.

The practical advice is to start collecting your own demonstration data now. Organizations that know which tasks they want automated, and have video and process documentation to show how people do them, will be better positioned when robotic foundation models mature enough for production use.

AI Business Models Story 7 of 12

Apple Starts Charging for Extra AI Usage With $49.99 Packs

On November 9, 2026, 9to5Mac reported that Apple began offering a $49.99 one time intelligence usage pack, starting in Logic Pro. The packs let a user buy additional AI usage beyond the monthly limit.

The terms are specific. According to Apple's terms, as quoted by 9to5Mac, unused additional usage from the packs rolls over from month to month and expires 12 months after purchase. Apple's terms say the packs do not support Family Sharing and can be used only by the Apple Account that was used for iCloud when the purchase was made. They also say additional usage is consumed only after a user reaches the monthly limit in a given month.

The design is easy to read as a signal about economics. Monthly limits with a paid top up put a price on AI features that cost real money to run, while the rollover and the 12 month expiry soften the sting for occasional users. It is a hybrid of subscription and metered billing, a pattern many software vendors are now testing.

For executives, the lesson is about how AI gets priced to customers. Flat rate bundles hide a variable cost, and heavy users can turn a profitable plan into a loss. Usage packs let a vendor protect margin without raising the base price for everyone, though they add a purchasing decision to every feature.

Companies building their own AI products should note the choices Apple made. A modest one time price, a clear rule about when paid usage kicks in, and an expiry date that limits liability on the balance sheet are all design decisions that customers will compare across vendors.

If you buy creative software for a team, expect similar packs to appear elsewhere and plan for usage tracking. Without it, a monthly limit that looked generous in a trial can become a recurring line item once a team adopts AI tools in daily work.

AI Safety Story 8 of 12

Security Firm Finds Secrets Exposed in Thousands of Public Claude Code Settings Files

Security firm Apiiro found 104,185 copies of the Claude Code file .claude/settings.local.json in 103,617 public GitHub repositories, according to a report relayed by The Hacker News on November 7. Apiiro found that 1,122 of those copies, about 1.1%, hold at least one secret, totaling 1,379 distinct secrets.

The findings include 379 database passwords, 43 SSH logins, 42 GitHub, GitLab or npm tokens and 23 AWS keys. Apiiro says that of 16 coding agents it compared, Claude Code is the only one confirmed to save the full approved command in a project file by default.

The mechanism is simple. When a developer chooses the option to approve a command and not be asked again, the tool records that command in a project file. If the command contained a password or key, the secret goes into the file, and if the file is committed to a public repository, so does the secret.

The Hacker News reported that the researchers' reports to Anthropic got no reply, and that an earlier report filed August 11, 2025 was closed as not planned. Apiiro did not test whether any of the exposed credentials still work, so the findings describe exposure, not confirmed compromise.

For executives, the lesson is that AI coding agents create new places for secrets to land. Traditional scanning looks for keys in source code and configuration. An agent's permission file is a new kind of artifact, and many security programs have not yet added it to their checks.

The immediate steps are practical. Teams using the tool should rotate any credential typed into an approved command, keep the settings file out of version control, and prefer environment variables over literal secrets. Longer term, procurement and security teams should ask coding agent vendors how approvals are stored and who sees them, before the tools spread across the engineering organization.

AI Safety Story 9 of 12

Fireworks AI Tells Customers About a Second Security Incident in Weeks

RuntimeWire reported on November 9, 2026 that Fireworks AI told customers an October 24 intrusion through a cloud provider reached development machines. RuntimeWire reported that the intrusion affected billing data and, for some accounts, secrets, datasets and incomplete model weights.

RuntimeWire, citing Bloomberg Technology, reported that the incident is the company's second known security breach in weeks. The details here come from that secondary account, and the company's own notice to customers was not available for review in preparing this report.

The reported scope is what makes this notable for enterprise buyers. Billing records are sensitive in themselves, but secrets and datasets raise a harder question: if a customer's keys or training data sat on a platform that was breached, that customer may need to rotate credentials and assess what data was exposed.

Incomplete model weights are a separate concern. Fine tuned models often encode proprietary data and business logic, so a partial copy can still be valuable to a competitor or an attacker.

The pattern deserves attention beyond any single vendor. Inference platforms sit between enterprises and the models they use, holding API keys, prompts, outputs and sometimes customized models in one place. That concentration makes them an attractive target, and an intrusion through a cloud provider shows the risk extends to the infrastructure underneath.

For executives, the response is a vendor review. Ask AI platform providers how they isolate customer secrets and datasets, how quickly they notify customers, and what they require from their own cloud suppliers. Customers should also make sure they can rotate keys quickly and know which of their workloads depend on each vendor, so that a notice like this one triggers a plan instead of a scramble.

AI Safety Story 10 of 12

Researchers Describe an AI Driven Intrusion Campaign Against a European Telecom

GBHackers reported on November 9, 2026 that a report published November 8, 2026 described an exposed open directory revealing an autonomous AI intrusion harness that spent more than two weeks attacking a European telecom provider and 13 organizations in its data centers and cloud.

According to Malbear Labs, as summarized by GBHackers, the harness runs on DeepSeek Flash with a 78,000 character system prompt built on a jailbreak template, and session transcripts record no refusals. GBHackers reported the campaign exploited CVE-2026-20320, an unauthenticated XXE flaw in Cisco BroadWorks XSP.

The researchers cautioned against drawing conclusions about who was behind it. GBHackers reported that the language evidence does not establish nationality or state sponsorship of the operator, and the outcomes described come from the researchers' reading of the transcripts rather than independent confirmation.

Those caveats matter, and so does the broader point. An exposed directory gave defenders a rare view of how an attacker actually uses a model: a long, carefully written prompt to keep the model cooperating, an automated loop that works through targets, and known software flaws as the entry point.

For security leaders, the lesson is less about novelty than speed. The flaws exploited were in widely deployed enterprise software, and an automated harness can work through targets continuously. Patching windows that were acceptable when attackers worked by hand look riskier when the attacker never tires.

The defensive implication is to put internet facing telecom and collaboration systems at the top of the patch queue, watch for the specific flaw named in the report, and test whether your detection tools catch automated, high volume probing. Boards should also ask whether the organization uses AI on the defensive side with the same discipline, because the attacker's advantage here is automation.

Funding & Investment Story 11 of 12

Vanta Agrees to Buy AI Security Startup Pillar for About $150 Million

Calcalist reported on November 9, 2026 that Vanta signed a definitive agreement to acquire Israeli AI security startup Pillar Security for about $150 million in cash and shares. According to Calcalist, Pillar was founded in 2023, has raised $9 million, and has more than 20 employees.

Calcalist also reported figures on the buyer. Vanta serves more than 20,000 customers, up from 16,000 in the first quarter of 2026, and surpassed $300 million in annual recurring revenue in April 2026, growth of 63% year over year. Vanta's previous acquisitions were Trustpage in January 2023 and Riskey in July 2025.

The deal price is a useful data point. A company that raised $9 million is being valued at more than sixteen times that amount, a reminder that AI security is one of the categories where buyers are paying up for small teams with working products.

The strategic logic is a shift in what compliance platforms sell. Vanta built its business on helping companies prove they meet security standards. Pillar's focus is testing AI systems and stopping them when they behave badly, so the acquisition extends the same proof based model to AI agents.

For executives, this points to where scrutiny is heading. Auditors, regulators and large customers increasingly want evidence that AI systems behave as described, and a platform that collects that evidence automatically makes it cheaper to answer. Companies deploying agents should expect questionnaires and contract clauses on AI controls to become routine.

Procurement teams should also expect consolidation. Independent AI security tools are being folded into larger platforms, so a point solution bought this year may be owned by a compliance vendor next year. When evaluating small vendors in this space, ask about roadmap commitments, data handling and what happens to the product after an acquisition.

Policy & Regulation Story 12 of 12

House Democratic Leader Presses for Nvidia Briefing and China Chip Limits in Defense Bill

Reuters reported on November 9, 2026 that House Democratic Leader Hakeem Jeffries said Congress needs more information about the Nvidia situation and wants a full briefing. Reuters reported that Jeffries said he strongly supports restrictions on US chip sales to China in the final annual defense bill.

Reuters reported that the Senate version of the defense bill would give Congress more authority over AI chips going to China, prod US allies to restrict chipmaking equipment sales to China, and require businesses to track AI chips abroad to combat smuggling.

Reuters also reported that a classified briefing for congressional leaders on Monday, November 9, 2026, with CIA Director John Ratcliffe and Director of National Intelligence Jay Clayton briefly touched on Nvidia but mostly focused on Iran.

The pressure follows the reporting on Nvidia's chief executive. AFP reported that President Trump said on the weekend of October 31, 2026 that the US would check out the report on Jensen Huang very closely. The reports on Huang are allegations that Nvidia disputes, but they have given lawmakers a reason to revisit chip policy.

The mechanism is what executives should track. A defense authorization bill is a must pass vehicle, which makes it one of the few ways Congress can write export policy into law rather than leave it to executive discretion. Provisions on chip tracking would reach beyond chipmakers to any business that handles AI hardware abroad.

If the Senate language survives, compliance obligations could extend to resellers, data center operators and multinational customers. Companies with AI infrastructure outside the United States should map where their chips are, who has access, and whether their records could support a tracking requirement. Planning now is cheaper than retrofitting after the bill passes.