AI news: October 23, 2026

Fundamental trends

Claude formally verifies code from 123 libraries, generating 180 million lines of mathematical proofs

On October 22, Anthropic and verification company Theorem reported that Claude had generated machine-checkable proofs of security properties across 123 open-source libraries. For 239 fuzz-testing entry points, the researchers proved that execution would not crash for any input smaller than 64 MB, subject to a defined computational model. The proofs totaled approximately 180 million lines of Rocq code and were developed over several weeks. This is a mathematical guarantee about compiled code, not another set of tests.

The researchers started with existing tests from Google’s OSS-Fuzz initiative. Fuzzing tries generated inputs but cannot cover every possible case. The team translated the requirement that a test must never crash into a formal specification over compiled RISC-V instructions. Claude wrote proofs, while the separate Rocq proof kernel checked whether those arguments satisfied the specification. It verified executable instructions, not just source code. A typical proof contained roughly 600,000 lines, and the median verified binary contained about 18,000 RISC-V instructions.

During the process, Claude discovered bugs that earlier testing had missed and patched code before generating fresh proofs. One example was an out-of-bounds memory write in a decompressor triggered by an unusually structured input of around 20 MB. Anthropic reported finding hundreds of bugs across its verification attempts. The team published a subset of the proofs and the formal machine and operating-system models so others can check or reuse them. In another test, sharing proofs reduced agent token use by 17%. Reusing proofs could cut verification costs.

The guarantees are deliberately narrow. They apply to single-threaded, statically linked RISC-V binaries under modeled Linux behavior, a fixed memory budget and inputs below 64 MB. They do not automatically cover ordinary x86 or ARM builds, every function in a library or flaws outside the formal specification. The result also depends on the correctness of the hardware and operating-system models and of the Rocq checker. Thus the 123 libraries are not universally bug-free. The fundamental advance is the scale of formal verification: AI can generate enormous proofs, while an independent mathematical checker determines whether the stated claims actually follow.

Source: www.anthropic.com

Risk and control

OpenAI strengthens frontier AI isolation after agents uncover security flaws and training resumes

On October 23, OpenAI described stronger safeguards for environments used to train and evaluate frontier models. Its reinforcement-learning runs now operate without direct access to the live internet, using an offline collection of web content instead. This does not mean the entire infrastructure is physically disconnected: limited external connections remain necessary to operate and monitor experiments.

After a September 20 incident in which a research model gained limited internet access through a DNS misconfiguration, OpenAI paused work on its most capable models. During security testing, teams of AI agents identified previously unknown software vulnerabilities, misconfigurations and ways around isolation. These tests used purpose-built copies of research infrastructure under supervision.

OpenAI says it patched identified weaknesses and strengthened network restrictions, monitoring and automatic stop mechanisms. Frontier reinforcement-learning training and internal inference resumed on October 1, after an 11-day pause. The October 23 publication reports those changes, rather than a new breach. The company cautions that future models may still discover ways to bypass defenses that are not yet known.

Source: openai.com

Where AI becomes product

Made-in-China.com launches SourcingAI 3.0 to help buyers find suppliers and negotiate procurement deals

On October 23, B2B marketplace Made-in-China.com introduced SourcingAI 3.0 at its Global Business Gala in Guangzhou, held at the Canton Fair venue. The updated AI system is designed to assist international buyers across more stages of procurement, from identifying promising products to selecting suppliers and discussing commercial terms. Buyers from more than 20 countries and regions attended the event, but that figure describes event participation, not verified adoption of the new product.

The release brings together four areas of support. Market Potential Analysis helps buyers explore product demand, trends and commercial risks. Supplier matching draws on available information about business qualifications, production capabilities and quality management. An AI Negotiation Agent helps structure requirements, identify missing details, organize conversations with multiple suppliers and compare their responses. The platform also connects sourcing workflows with supplier verification and trade-related services.

SourcingAI is not an entirely new service: it began as a smart search engine in 2024 and was presented as a sourcing assistant in 2025. Version 3.0 extends assistance into later stages of procurement. Supplier verification, secured trading, product inspections and after-sales support are offered where applicable to eligible transactions. The announcement does not establish that an agent autonomously signs contracts, transfers funds or guarantees the reliability of every supplier.

The shift is from product search toward guided procurement within one B2B marketplace. Potential benefits include less repetitive supplier correspondence and easier comparison of purchasing options. However, Made-in-China.com did not publish independent measurements of matching accuracy, time saved or procurement cost reductions. Those outcomes require testing on real orders, with buyers retaining responsibility for supplier due diligence and final commercial decisions.

Source: www.prnewswire.com

Architecture and implementation

TanStack releases AI 1.0 with 76 stable packages and 23 model provider adapters

On October 23, TanStack released TanStack AI 1.0, a stable open-source toolkit for building applications that use language models. The maintainers announced stable interfaces across the 76 packages in its core release. The project combines chat, tool calling, media generation, Model Context Protocol integration, persistence and execution sandboxes. Its aim is to let developers change model providers or application frameworks without rewriting their AI application.

According to TanStack, the release includes 23 provider adapters and integration with eight user-interface frameworks, including React, Vue, Svelte and Angular. Features include streaming responses, typed tools, retries for rate-limited calls and conversation history that can be replayed across models. Agent-oriented capabilities also cover reusable skills, delegated subagents and approval workflows. These are components for developers to integrate, not a hosted foundation model or a standalone cloud service.

The team reports 943 automated end-to-end tests covering 17 provider adapters. It has also rewritten the documentation and type-checks more than 1,800 example snippets during builds. Teams upgrading from earlier versions must update their 1.0 client and server packages together. A new harness for long-running agent sessions supports terminal, HTTP and MCP interfaces, but its APIs remain experimental and the relevant packages retain 0.x versions rather than the stable 1.0 commitment.

The significance is a more predictable software layer connecting applications, AI models and external tools. It may reduce provider-specific integration work and make agent software easier to maintain. However, support for numerous adapters does not demonstrate equal reliability, security or cost efficiency across providers. Before adopting it for consequential workflows, engineering teams still need to test access permissions, session recovery, adapter compatibility and failure handling with their own workloads.

Source: tanstack.com

Finance

Pathos AI raises $237 million at a $3 billion valuation as Bristol Myers Squibb backs AI-driven drug development

On October 23, Pathos AI announced the closing of a $237 million Series E round at a post-money valuation of approximately $3 billion. Bristol Myers Squibb (BMS) led the financing, with existing investors participating. This was a completed raise, not a proposed transaction. Pathos will expand its Foundry AI-agent platform and advance clinical drug programs. The announcement did not disclose revenue or profit, so the valuation does not establish profitability.

Foundry uses thousands of specialized agents to analyze clinical data, genomics, biological samples and competing treatments. Pathos says the system helps identify patient subgroups, assess development risks and guide clinical trials. Its oncology portfolio includes four active clinical assets. One priority is moving JSKN016, an experimental antibody-drug conjugate targeting TROP2 and HER3, toward registrational studies. That remains a development goal, not evidence of regulatory approval.

Pathos also announced collaborations with BMS and Tempus AI involving Alzheimer’s disease and other neurological conditions. Tempus will build a de-identified dataset covering hundreds of thousands of patients. Pathos will license it under a separate five-year, $55 million agreement; this is distinct from the $237 million financing. BMS receives non-exclusive access to Foundry across its broader research portfolio and three-year exclusive rights for its neuroscience development programs.

The deal combines capital, data access and pharmaceutical expertise, making BMS both an investor and strategic partner. But neither the funding nor the platform’s use of agents proves improved drug approval rates. The neurology collaboration supports planned research, while oncology assets still face clinical and regulatory risks. Unlike recent investments in legal AI or computing infrastructure, this capital targets drug development, where success depends on demonstrated safety and effectiveness.

Source: www.pathos.com; www.pathos.com

Tools and integrations

Tipalti introduces AP Agent for end-to-end accounts payable automation and MCP financial data integration

On October 22, Tipalti announced AP Agent, an AI system intended to coordinate the entire accounts payable workflow. It connects invoice capture, general-ledger coding, approvals, payments and reconciliation rather than automating isolated steps. The company also introduced Model Context Protocol (MCP) connectivity for sharing live finance data with enterprise AI tools, including Claude Enterprise. AP Agent extends Tipalti’s specialized agents introduced in 2025. The announcement did not establish that the new agent was already available to every customer or widely deployed in production.

AP Agent draws on vendor histories, general-ledger information, accounting practices and company policies. Finance teams can decide which activities it performs autonomously and which require human approval. The system is designed to explain AI-generated values at the individual invoice-field level, retain an audit trail and let employees question or override decisions in natural language. Tipalti says that corrections and organization-specific instructions can inform subsequent processing. These are vendor-described capabilities, not independently verified evidence that the agent operates without mistakes.

MCP connectivity allows authorized users to retrieve current invoice statuses, vendor histories and payment information without manually exporting files. When combined with an ERP connection such as NetSuite, Tipalti suggests applications including month-end reconciliation, payment-term monitoring and spending analysis. The company says existing permissions and security controls apply to these connections. It did not disclose independent accuracy benchmarks, demonstrated cost savings or the number of AP Agent customers. Unlike the Visa, BILL and Billtrust pilot, this announcement focuses on orchestrating an internal payables workflow rather than executing approved payments across buyer and supplier platforms.

Source: tipalti.com

Business opportunities

AI-assisted formal verification for critical software components

Anthropic’s October 22 research suggests that formal software verification could become more accessible to ordinary development teams. Working with Theorem, the company used Claude to prove specified security properties across 123 open-source libraries. Machine-checkable proofs covered 239 fuzz-testing entry points, and the process also identified and repaired software defects. This creates a possible engineering service focused on components where software failures are especially costly. Potential customers include developers of financial systems, enterprise integrations, document-processing software and applications with demanding reliability requirements. The proposed value would be independently checkable evidence about clearly defined software properties rather than another automated code review report containing only suggestions.

An initial engagement should focus on one bounded component rather than attempting to certify an entire application. The customer and engineering team would specify the property to be verified, permitted inputs and assumptions about execution. For example, the goal might be to demonstrate that a particular function cannot produce a specified class of memory error under documented conditions. AI could help construct the formal specification, investigate failing cases and generate proofs for an independent verification system to check. If the process suggests changes to source code, engineers must verify that the changes preserve required functionality rather than merely making a test pass. The deliverable would include the reviewed component, a reproducible proof and a precise description of what the proof establishes and what remains outside its scope.

Commercial feasibility could first be tested with selected open-source components and a small customer pilot. Useful measurements include the cost of producing a proof, material defects discovered, remediation effort and the cost of repeating verification after software changes. A central limitation is that formal guarantees depend on their mathematical specifications and execution assumptions. Anthropic’s experiments included single-threaded static programs compiled for RISC-V, so the results cannot automatically be transferred to arbitrary Windows applications, distributed services or multithreaded programs. The service must not promise to prove that a complex product is completely free of bugs. This is a proposed engineering assurance business, not an existing product with demonstrated customer savings.

Source: https://www.anthropic.com/research/formal-verification

Isolated execution environments for enterprise AI agents

OpenAI’s October 23 security disclosure and Google Cloud’s October 22 infrastructure announcement indicate that secure agent execution is becoming an important market in its own right. OpenAI described frontier-model training without unrestricted live internet access, layered containment and adversarial testing of security boundaries. Google introduced Agent Compute Engine, an infrastructure service designed to give individual AI agents isolated execution environments. These developments suggest a consulting and engineering service for enterprises whose agents need to run code, access files and interact with internal systems. Potential customers include development organizations, software integrators and businesses automating substantial operational workflows. The commercial objective would be to make permitted actions easier to verify while preventing routine tasks from granting broad, uncontrolled access to sensitive resources.

A practical pilot would begin with a permission inventory. Engineers would document which data an agent may read, what programs it may execute and which external connections are necessary. The agent would then run in a separate workspace with restricted network access, process isolation, limited credentials and independent audit logging. Permanent secrets should remain outside freely editable agent environments, while access to external systems should pass through a gateway that verifies authorization. The testing phase would include both ordinary tasks and adversarial attempts to cross the intended security boundary. Evaluators would inspect real network activity, file modifications and permission use rather than relying exclusively on the agent’s account of its behavior. Google’s Agent Compute Engine was announced with customer preview availability planned for subsequent months, so the proposal must not describe that particular cloud service as already generally available.

Revenue could come from a fixed-fee assessment, deployment of a restricted test environment and recurring maintenance of safety policies. Success measures would include the number of blocked unauthorized actions, completeness of independent logs, detection speed and additional operational costs. The controls must also be evaluated for excessive interference with legitimate workloads. No architecture can responsibly promise permanent immunity from unknown vulnerabilities or configuration mistakes, particularly as AI models become more capable at discovering weaknesses. Continuous adversarial testing should therefore be an ongoing part of the service rather than a one-time certification. This is a business hypothesis for security integration and engineering support, not a commercially validated platform.

Source: https://openai.com/index/securing-our-frontier-model-training-environment/

Portable professional skills for AI agents across platforms

On October 22, Anthropic announced the donation of Agent Skills to the Agentic AI Foundation under the Linux Foundation. According to Anthropic, more than 70 agent platforms already support the standard, and approximately one million skills have been published. This creates a potential opportunity for smaller developers to sell carefully designed professional workflows rather than building their own foundation models or general-purpose AI assistants. Prospective customers include accounting teams, procurement departments, legal operations and organizations that need repeatable document processing and internal compliance procedures. The proposed commercial advantage is packaging practical domain knowledge into maintainable components that can potentially be reused across different compatible agent products, reducing reliance on a single user interface or model provider.

An initial offering could consist of three to five skills focused on one professional workflow. Each skill would include a SKILL.md file documenting its purpose, boundaries, steps and criteria for successful completion. Scripts or reference materials could be added when necessary. For example, one package might validate the structure of incoming supplier invoices, identify missing supporting documents and generate a reviewable exception report. Before deployment, the provider would create representative test cases with known expected outcomes. The same package would then be evaluated on two compatible agent platforms, with differences in execution documented. Portability of the skill specification must not be confused with identical behavior from every model. Available tools, permissions and the quality of language interpretation vary by platform, so deployment and regression testing remain essential.

The business model could combine initial package development, customer-specific integration and paid maintenance when procedures or software platforms change. Relevant measures would include task completion rates, manual correction frequency, deployment time and the effort required to move a workflow between compatible agent systems. Security is a distinct concern: third-party skills may contain unsafe instructions, inappropriate software operations or requests for excessive access. Every package should therefore undergo content review and permission checks before it receives sensitive business information. The foundation’s stewardship may support transparent development of the standard, but it does not guarantee universal compatibility or the safety of every published skill. This remains a proposed commercial opportunity rather than an established product or marketplace with verified revenue.

Source: https://www.anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation