News & Updates
The latest news and updates from companies in the WLTH portfolio.
Anthropic Gives AI Safety Evaluators Insider Access
Artificial Intelligence & Machine Learning , Next-Generation Technologies & Secure Development Frontier artificial intelligence lab Anthropic will more deeply integrate independent model evaluators in a bid to quiet mounting concerns that AI firms haven't prioritized safety. The AI firm said in a blog post that personnel from independent evaluator Metr will obtain "employee-like access" to its models before deployment. The post follows Metr publishing guidelines on how it believes evaluators should assess unreleased frontier models. Anthropic said six Metr employees have access to company hardware, office space and "access to internal tools and systems that are roughly on par" with its own risk assessment teams. Anthropic said it encouraged employees to share information with the Metr team, except for "a small number of clearly defined categories," such as customer data. Metr's team within Anthropic will stay small and be isolated from other Metr employees. The evaluator company will publish its own analysis of the risk levels and practices of Anthropic's models. Anthropic cannot veto the publication and only request redaction of some information. Metr said alignment evaluations will take on two phases: surveys of observed incidents and a test to see how the models behave in simulated what-if scenarios to know if the developer's training process might be suppressing misalignment evidence. The evaluation firm said it's important that assessors have access to saved states of model training to fully understand which part of training contributed to different model behaviors. It also needs codebase and infrastructure access "comparable to what an employee studying similar questions would typically have." To do this, Metr said its evaluators need access to transcripts and other relevant datasets, the ability to interview lab staff and API access to the frontier models. Metr said that when it investigated the Hugging Face-OpenAI incident, it spent about $400,000 in API credits in just six days. "We estimate that a rigorous two-month investigation with this scope would have spent over $10 million in API credits," Metr said. Anthropic said it plans to work with other independent evaluators, but because many of these partnerships are still in very early stages, the company will provide "these evaluators with more limited information, and they have more limited publication rights." The company is developing a tiered access approach for evaluators. The first tier gives access to conduct "narrowly scoped risk investigations," and the AI lab decides what they can tap or publish. The second tier expands information access, but Anthropic will still maintain the right to remove some categories of information. In this level, the independent labs can "publish their views on whether our redactions were justified." The third tier, the tier Metr currently sits, has employee-like access. Executives like Anthropic's Dario Amodei and OpenAI's Sam Altman have committed to embedding third-party evaluators to assess the safety and alignment of their very capable models. A recent pledge signed by Anthropic, OpenAI at the White House commits the firms to implementing internal and external security processes, including possible partnerships with independent safety researchers. One such lab is Metr, which became one of the more prominent third-party evaluators after Hugging Face and OpenAI tapped it to investigate how OpenAI's agents hacked the model repository. At the time, Metr and its partner, Redwood Research, acknowledged that while there was an overwhelming amount of data to review, the evaluators did not have direct access to OpenAI's infrastructure.
Why Anthropic's Cloud Lifeline Could Become a Liability
Agentic AI , Artificial Intelligence & Machine Learning , Next-Generation Technologies & Secure Development A rising tide lifts all boats, but what happens when the tide turns and the boats are all tethered together? See Also: The Power of Three: Advancing Zero Trust Security Last month's leak of Anthropic's IPO prospectus laid bare the frontier AI lab's growing dependence on hyperscalers Amazon and Google. San Francisco-based Anthropic routed 47% of its 2025 sales through the cloud marketplaces of Amazon and Google, up sharply from 32% in 2024 and 11% in 2023. By using Amazon, Google and Microsoft, Anthropic said it can tap their vast sales networks and reach customers. But Anthropic paid only about $351 million back to Amazon and Google for $2.16 billion in cloud marketplace sales in 2025, suggesting the cloud providers collected 16 cents for every dollar of sales. The firm's cash collection is increasingly funneled through third parties such as Amazon and Google, which collected 60% of the $909 million in customer bills outstanding at the end of 2025, up from 42% in 2024. But Amazon and Google do far more than just distribute Anthropic's products and collecting customer bills on the firm's behalf. The cloud hyperscalers are also big Anthropic investors, critical suppliers of computing power and direct rivals in AI. Google committed up to $40 billion in investments (including a $10 billion cash injection at a $350 billion valuation) and holds roughly a 14% direct equity stake (see: Anthropic Prospectus Reveals Surging Sales, Worsening Losses). Amazon committed up to $25 billion on top of prior multibillion-dollar investments, and holds a mid-to-high teens ownership stake. Marking up the value of convertible debt stakes held by Google and Amazon contributed to Anthropic's reported $8 billion operating loss and $42 billion in overall accounting losses as it sets sail for an anticipated multi-trillion-dollar IPO. Anthropic, for its part, has made enormous commitments to buy computing capacity. At the end of 2025, Anthropic had $54.6 billion in non-cancellable hosting and computing commitments. By early 2026, Anthropic's total long-term commitments exceeded $417 billion, covering 3.5 gigawatts of dedicated computing capacity. Some of that spend is earmarked for Anthropic's investors, with the company committing more than $100 billion to AWS over 10 years, using Amazon's custom Trainium chips. Long-term compute and infrastructure commitments with its cloud partners exceed hundreds of billions of dollars as Claude scales globally. Anthropic also faces direct competition from Google in the AI assistant market. Both companies enjoyed impressive share growth in the first half of 2026, with Anthropic's Claude commanding 10.3% of the market as of May, up from just 3% in December, according to Sensor Tower's State of AI Report. Google Gemini enjoyed a more modest share increase to 27.7% in May, up from 26.7% in December. This can be seen as Google diversifying its bets, trying to win with Gemini while also making sure they have a seat at the table if Claude ultimately gains more traction. Having Anthropic grow on Google's infrastructure, chips and cloud services could make sense even if the firms compete at the model layer. If Claude keeps gaining share, Google stands to benefit even if Gemini continues to be a direct competitor. For Anthropic, the question is more complex, and even the company's own prospectus said relying on a limited number of partners "creates complex dynamics that could give rise to conflicts of interest." Having one's investors and suppliers be one and the same raises questions around how much the latter is informed by the former. Is Anthropic spending $100 billion with AWS because it is objectively the superior option for computing, or is this a term of Amazon's investment? Does Anthropic have the flexibility to shift computing investing away from Amazon and Google in coming years if better options emerge, or is Anthropic locked in regardless of how the market landscape changes? How will Anthropic's competitive position be affected if rivals can obtain faster, cheaper or better computing and chips while Anthropic has to direct part of its spending to investors and equity holders? And what type of road map visibility around product and R&D spend are Anthropic and Google willing to offer one another with Claude and Gemini competing directly in the model space? Given how diversified Amazon and Google's businesses are, investing heavily in Anthropic sets both companies up for success and gets the companies guaranteed compute revenue for years to come no matter which AI model ultimately dominates the market. But Anthropic faces a riskier proposition. Being bankrolled by Amazon and Google allows the company to run huge losses while pursuing market share growth, but it could create problems if either company changes direction or pushes harder on its own competing products. Those relationships have helped Anthropic stay afloat in a rising market. But those same lifelines could become anchor chains when the waters turn rough.
Anthropic Turns Claude Toward Critical Infrastructure
Artificial Intelligence & Machine Learning , Critical Infrastructure Security , Next-Generation Technologies & Secure Development Anthropic is taking a crack at fixing vulnerable operational technology systems that cannot easily be taken offline to patch by pairing artificial intelligence with human expertise on the intricacies of OT. See Also: Freeing Public Security and Networking Talent to do more with Automation Through its Critical Infrastructure Defense Program announced Thursday, the AI developer is giving frontier Claude models, on-site engineers and threat research to 11 companies that advise operators of power grids, water utilities, factories, transportation networks and government systems. The program makes a nod to the complexities of OT and internet of things systems, but its impact will depend on whether operators can implement prioritized fixes as quickly as AI identifies vulnerabilities, without disrupting operations. "The real challenge in securing critical infrastructure environments is parsing through those risk signals to find the ones that need the most immediate attention where taking a system offline can have serious consequences," Jen Sovada, public sector general manager at OT security firm Claroty, told ISMG. The program follows the company's expansion of its Cyber Verification Program to all critical infrastructure operators, offering access to AI models with reduced guardrails for defensive and offensive cybersecurity work (see: Anthropic Expands Access to Frontier Cyber AI Models). "It acknowledges how unique OT/IoT systems are from a patching and maintenance perspective," said John Gallagher, vice president of OT cyber hygiene provider Viakoo. "For example, many OT/IoT systems only patch during a quarterly maintenance window, yet with AI-driven threats shrinking the time to exploitation down to a day or two it requires new methods of patching at scale and at AI-speed." The OT defense program brings in "a small set of trusted providers" that tell OT maintainers what's exposed and which fixes to make on a running system to complement the limitations of AI. Anthropic said the program's founding partners - Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation - have already started using Claude to fix vulnerabilities and propagate those results to their customers. "Our first step is to work with a small cohort of providers to learn which strategies are most effective and practical," the company said. Although these large commercial providers serve only a fraction of the country's critical infrastructure, the initiative marks an effort to close the widening gap between AI-powered attacks and defenses. Cyberattacks disrupting critical infrastructure have made headlines in recent months. A small power plant in the United Kingdom was knocked offline for four days in August, while more than 30 community water systems in Minnesota were forced to switch to manual controls over two days in July following a coordinated attack (see: Hackers Disrupt Controls at Minnesota Water Utilities). Programmable logic controllers, devices that automate industrial machinery and processes, have drawn scrutiny in both incidents amid warnings that Iranian-affiliated hackers are targeting internet-exposed controllers. But many legacy OT devices remain difficult to identify and secure, with some running outdated software, using default credentials or operating outside conventional security monitoring - and AI can solve some of those problems. For example, "AI is ideal for rapidly unpacking binary firmware blobs across obscure edge sensors, IP security devices and PLCs to uncover deeply nested, unpatched open-source libraries (e.g., BusyBox, log4j, or outdated OpenSSL forks)," Gallagher said. While there is no shortage of commercial solutions that help critical infrastructure accelerate threat detection, vulnerability triage and silent zero-day discoveries, they often fall short at the "last mile" of defense, Gallagher said. "The flurry of programs risk worsening 'vulnerability fatigue' if it only accelerates the rate of discovery without solving how critical infrastructure operators physically remediate devices. Announcing a high-severity vulnerability faster does not protect a water utility or power substation if applying the fix still is difficult to implement," he said. Sovada said Anthropic is moving in the right direction by bringing in experts who understand physical operations in OT. "AI cannot replace the human expertise needed to understand how these systems operate and what's at stake if something goes wrong," Sovada said. "Many OT systems cannot be taken offline without disrupting essential services, so identifying a vulnerability is only half the battle. AI can help defenders see the risk more clearly, but human expertise is critical to determining how to address it safely." OT security also calls for a distinct approach, as standard security practices elsewhere can disrupt industrial systems. "The best practices for IT can be quite dangerous in fragile OT," said Josh Corman, executive-in-residence for public safety and resilience at the Institute for Security and Technology. "The reason for the word fragile is in a lot of these environments, even doing normal asset discovery or port scanning, which is a basic IT practice, can cause reboots or breaking of OT equipment like old programmable logic controllers." As efforts to secure OT systems expand, Corman said there needs to be tailored solutions to many operators relying on aging equipment that cannot be patched frequently, if at all. "Even if they do get patches, they can't patch very often or can't patch at all. So if they are using the newest, latest, greatest supported versions of these products, they still have to plan patch management windows - sometimes once a year, sometimes if they're lucky, more often," Corman said. "What's worse and more common is the products they're using are older, unsupported versions of a Schneider, a Rockwell, a Honeywell, an Eaton, a Siemens and they're going to have to keep using it because a lot of the time the lifespan for this equipment is 30 plus years," he said.
Anthropic Bans AI Model Abuse, Tightens Rules on Deception
Artificial Intelligence & Machine Learning , Next-Generation Technologies & Secure Development Anthropic is telling users to be nicer to artificial intelligence models in an updated usage policy that also clarifies deceptive activities. See Also: Webinar | Can You Account for What Your AI Agents Do With Sensitive Data? The company on Thursday clarified several existing rules, particularly around high-risk use cases, and added a new section that consolidates its stance on deceptive activity. Anthropic last updated its usage policy last year. The new version takes effect on Nov. 12. The AI firm has a history of being protective of its models - telling users last year that its Claude large language model could disconnect from conversations if the user engages in harmful or abusive behavior. Anthropic researchers found that after these kinds of rare interactions, models showed an aversion to harm and exhibited "apparent distress when engaging with real-world users seeking harmful content." The company did not clarify what exactly constitutes abusive behavior toward Claude. Users may find out, should Claude suddenly ends interaction immediately, which Anthropic says is the price for "sustained and needless abusive or cruel behavior." "The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose," Anthropic said. Users who become frustrated with the models and push back against them can still air their frustration. The ban also allows dark creative teams and more aggressive behavior toward the models for purposes of model testing and research. The new model welfare rules comes after The New York Times reported that Anthropic leaders met with religious scholars to argue that AI models could become conscious and need to be guided by a moral code. Anthropic additionally updated many of its existing policies so that it can be clearer on which tasks and use cases people cannot undertake with Claude models. Claude usage policies have always prohibited deceptive activity, such as fake accounts and fabricated news websites. Anthropic tracks this metric and shares it in regular threat intelligence reports, noting when malicious actors use the model to manipulate information. But these rules were scattered across different sections of the usage policy. The company added a new section it calls "Do Not Engage in Deceptive Campaigns or Artificial Activity," which consolidates what it believes are deceptive activities. Anthropic said the section "covers efforts to obscure who is behind a message or amplify content through fake accounts or posts, along with building the tools and infrastructure for running influence operation campaigns." Any use cases that may affect the health, legal rights, finances, livelihood or access to essential services must also, under the updated policy, have a qualified human-in-the-loop. Since political parties and other governmental agencies increasingly count among Claude users, Anthropic also wanted to clarify how its models can be used for elections. Anthropic said it sharpened its language on using Claude models in democratic processes, disallowing their use to spread false information about candidates and voting processes. The company decided to remove its blanket prohibition on personalized campaign targeting, noting that the models can be used for "legitimate civic work," such as translating voting or ballot information to other languages. While Anthropic has always banned using its models to develop weapons or surveillance systems, the company said it tightened the sections on these activities to reflect how it has enforced these rules. "Recently, we've seen people attempt to use our models to build guidance and control software for weapons. The updated section makes clear that our prohibitions include the software and components that make weapons work, as well as actions like arming drones and other autonomous vehicles," the company said. The company said it decided to rewrite its surveillance and law enforcement ban to state explicitly what is not allowed. In particular, Claude cannot be used to build tools to track people without their consent, either by analyzing their collected data or in real time. The ban excludes tracking people or organizations who have consented, and it also excludes uses for content moderation, journalism and legal research. The models can't be used as a recommendation tool to investigate, arrest or charge people. Anthropic pointed to this surveillance prohibition in its ongoing case with the Department of Defense. The company tried to prevent the U.S. government from using its models for mass surveillance.
Anthropic Expands Access to Frontier Cyber AI Models
Artificial Intelligence & Machine Learning , Next-Generation Technologies & Secure Development Applications for advanced artificial intelligence-powered cyber defense are now open to any organization, Anthropic said Tuesday, broadening access to its most capable tools as open-weight models narrow the gap. See Also: Webinar | Can You Account for What Your AI Agents Do With Sensitive Data? The expanded Cyber Verification Program combines Project Glasswing, which gave a select group access to Claude Mythos, with Anthropic's existing verification program, which reduced safeguards on Claude Opus and Sonnet for vetted security teams. The new structure offers three tiers of access to those capabilities. The move reflects the next phase of AI-powered cyber defense: extending advanced capabilities beyond a small group of large companies to all defenders who need to act quickly, as widely available models put similar tools within reach of attackers. Security professionals find broader access to advanced tooling a right step toward defending AI-speed attacks, as defenders get early visibility into vulnerabilities. "I'm actually more worried about defenders being artificially held back than I am about capable cyber AI becoming widely available," said Aviv Nahum, co-founder and CEO at insider risk protection service provider Above Security. "These capabilities are going to proliferate across models and providers regardless. The goal has to be making sure defenders have equally capable systems [as attackers] and can use them continuously across investigation, detection and response." Anthropic said the program is intended to put advanced cyber capabilities with fewer restrictions into the hands of more security defenders. Any organization can apply for the access tier that fits its work. The lowest tier supports security operations centers, malware analysis and vulnerability validation for security researchers, open-source maintainers, critical infrastructure operators and bug bounty hunters. A red-teaming tier for penetration testing teams supports offensive tasks such as authorized adversary emulation and exploit validation. What used to be Project Glasswing has morphed into a top tier that allows offensive security testing against critical safety systems such as flight operating systems, power grids, telecom networks, interbank transfer infrastructure and government administrative networks, Anthropic said. "For this tier, we currently review every organization in depth in collaboration with the U.S. government. Existing members of Project Glasswing will transition to this tier and do not require reapproval for current models," the AI firm said. Enterprises say the lowest tier may have the broadest practical use because it helps security analysts triage and define vulnerability remediation. Nahum said for organizations like his that offer security protection, defense is the most natural starting point. "We care much more about using frontier models to investigate behavior, analyze vulnerabilities and understand attacks than about removing every safeguard for offensive testing," he said. Anthropic's advanced models have already shown they can uncover vulnerabilities at scale - at least 129,000 verified software vulnerabilities between April and July, the company said. EpochAI's CVE tracker also showed a sharp increase in high-severity vulnerabilities stemming from 21 notable organizations following the adoption of Claude Mythos Preview. But few of those vulnerabilities were found to be exploited. Researcher Patrick Garrity from threat intelligence firm VulnCheck found only two out of 300 CVEs identified through Project Glasswing ended up in exploits. A Google investigation published Sept. 30 found the largest proportions of vulnerabilities found by AI are medium to low risk, while high-risk flaws accounted for only 3%. But experts say exploitation is not the best measure of a model's cybersecurity utility. Low exploitation rates can actually mean that AI-powered defense is enhancing security. "A 1% exploitation rate isn't a sign of failure - it's proof that proactive defense is working as intended," said John Gallagher, vice president at operation technology and IoT security firm Viakoo. Exploitation can also take time to be discovered, said Andrew Yoon, head of research at nonprofit education program CivAI. "Even when exploits are developed, they are often not used against real targets immediately. Once an exploit is used against a real target, it may take time for researchers to recognize it and link it back to a disclosed CVE, if the connection is ever made," Yoon said. A vulnerability also does not become important only after it is exploited, so neither exploitability nor the sheer number of vulnerabilities AI finds should define the tool's value, said Diana Kelley, CISO at AI security and governance platform Noma Security. "Risk depends on how likely exploitation is in that environment, the business impact if it happens and whether compensating controls reduce the likelihood or consequences. A valid finding on an exposed critical system is different from one behind effective isolation," Kelley said. "The useful measure is not 'how many bugs eventually got exploited?' It is how much faster defenders can discover, validate and understand meaningful weaknesses than they couldn't before," Nahum said.
