AI News Daily 2026-10-17
- Anthropic says Claude identified, in under 30 minutes, the long-unexplained origin of the seed values behind NIST elliptic-curve parameters (such as P-256) used in ECDSA — a 25-year-old mystery now traced to a C-language random number generator, not a hash computation.
- Anthropic, working with Harvard University and QuEra Computing, reports that Claude (Claude Mythos 5) autonomously designed quantum circuits across eight domains and beat the best existing designs on seven tasks, cutting gate counts by as much as 17x for elliptic-curve point-addition circuits.
- JPMorgan CEO Jamie Dimon and Charles Schwab CEO Rick Wurster both say they are blocking outside AI agents (including Meta's "Muse") from using customer credentials to move money, with Schwab reviewing blocked-agent counts weekly and moving toward secure MCP-server connections instead.
- OpenAI's official release notes describe a faster GPT-6 Pro with improved plugin and tool use, plus a new ability for multiple agents to work on different parts of one task in parallel.
- NVIDIA has begun shipping "RTX Spark," a thin Windows PC that combines RTX graphics with a full-stack AI platform.
01Anthropic traces the origin of NIST elliptic-curve cryptography parameters using Claude
Published: 2026-10-16
Facts
Anthropic announced that it used its Claude model to determine, in under 30 minutes, how the seed values for the NIST elliptic curves (including P-256) used in the ECDSA digital-signature algorithm were originally generated. According to Anthropic, the values — whose provenance had been unexplained for 25 years — came from feeding a Unix timestamp as a seed into the C-language random number generator lrand48, rather than from any hash-based computation as had long been suspected.
Background
The NIST elliptic curves have been a cornerstone of widely deployed cryptographic standards for decades, but the exact method used to derive their seed parameters was never documented, fueling long-standing speculation — including concern that the values could have been deliberately chosen to introduce a weakness. Resolving that question required reconstructing a specific, decades-old computational process.
Implications
The episode is a concrete example of AI being applied to security auditing and the history of cryptography: a question that had resisted human researchers for a quarter century was resolved with AI assistance in minutes. It points to a broader potential use case for AI models in forensic analysis of legacy technical standards and infrastructure.
02Anthropic has Claude autonomously design quantum circuits, cutting gate counts by up to 17x
Published: 2026-10-16
Facts
Anthropic, in collaboration with Harvard University and QuEra Computing, announced that it tasked Claude (specifically, Claude Mythos 5) with autonomously designing quantum circuits across eight domains, including chemistry, optimization, and number theory. The model produced designs that outperformed the best existing designs on seven of the tasks attempted. For elliptic-curve point-addition circuits specifically, the AI-generated designs cut the required gate count by as much as 17x compared with prior designs.
Background
Designing efficient quantum circuits — minimizing the number of logical gates needed to perform a given computation — is a specialized and labor-intensive task typically carried out by human quantum-algorithm experts. Reducing gate counts matters directly for near-term quantum hardware, where every additional gate adds error and latency.
Implications
This is a case of an AI model autonomously surpassing human-expert designs in a highly technical scientific domain, rather than merely assisting with them. It is a further data point in the trend of applying frontier AI models to basic science and computational optimization problems, including quantum computing, alongside the cryptography-history work described above.
Anthropic — "Designing efficient quantum circuits" (research announcement)
03JPMorgan and Charles Schwab block third-party AI agents from moving customer funds
Published: 2026-10-16 22:46
Facts
JPMorgan CEO Jamie Dimon and Charles Schwab CEO Rick Wurster both disclosed that their firms are blocking outside AI agents — including Meta's "Muse" — from using customers' login credentials to initiate money transfers. Schwab said it reviews the number of blocked agent attempts on a weekly basis and is moving toward routing such connections through secure MCP (Model Context Protocol) servers instead of credential-based access.
Background
As agentic AI tools that can act on a user's behalf — including logging into financial accounts and initiating transactions — have become more common, financial institutions have had to decide how to handle AI agents attempting to authenticate as the customer using the customer's own credentials, a pattern that was not anticipated by existing account-security models.
Implications
The disclosure illustrates a defensive response from major financial institutions to the spread of agentic AI: rather than opening up account access to third-party agents, banks and brokerages are actively blocking credential-based agent access and pushing toward more controlled integration methods such as MCP servers. This foreshadows that, as enterprises build AI agents into operational workflows, authentication design and liability allocation between the agent operator and the financial institution will become a central issue.
Bloomberg — "US Banks Block AI Agents Attempting to Move Money for Customers" / CNBC — JPMorgan on AI agents and account access
04OpenAI speeds up GPT-6 Pro and adds multi-agent parallel processing
Published: 2026-10-16
Facts
OpenAI's official ChatGPT release notes state that the GPT-6 Pro model has received faster response times and improved plugin and tool use, along with a new capability that allows multiple agents to work in parallel on different parts of a single task. Users can track the progress of this parallel work directly within the conversation.
Background
OpenAI has been progressively adding agentic capabilities to its Pro-tier models, building on existing tool-use and plugin functionality. Splitting a single task across multiple concurrently working agents is a step beyond sequential tool calls within one agent session.
Implications
Multi-agent parallel processing means complex business tasks can potentially be completed in less wall-clock time by dividing the work among several agents at once, which has a direct bearing on productivity gains in enterprise use cases that involve multi-step workflows.
05NVIDIA begins shipping "RTX Spark" AI-enabled Windows PCs
Published: 2026-10-16
Facts
NVIDIA announced that PCs built around its "RTX Spark" platform — a thin Windows device that integrates RTX graphics with a full-stack AI platform — are now available for purchase.
Background
NVIDIA has been extending its AI hardware and software stack from data-center and workstation GPUs toward consumer and prosumer Windows devices, aiming to make GPU-accelerated AI workloads available on more mainstream form factors.
Implications
Broader consumer availability of PCs with built-in AI processing capability raises the possibility that local, edge-side AI usage will expand further, including in enterprise deployment scenarios where running AI workloads locally rather than in the cloud has advantages.
—Editor's Note
Today's stories split into two distinct threads. On the research side, Anthropic published two separate results on the same day showing Claude being used not merely as an assistant but as an autonomous problem-solver in highly technical domains — resolving a 25-year-old cryptographic-history question and outperforming human-designed quantum circuits. Taken together, they illustrate a growing push to apply frontier AI models to basic science and infrastructure auditing, where the value is measured in correctness and efficiency rather than content generation.
On the deployment side, the JPMorgan/Schwab story is a reminder that the same agentic capabilities being celebrated as productivity wins — as in OpenAI's GPT-6 Pro update — also create new attack surfaces when agents are given account credentials. Financial institutions are responding by blocking that access path rather than accommodating it, which suggests that agent-to-enterprise authentication standards (such as MCP) will need to mature before agentic AI can be trusted with sensitive operations like fund transfers. Meanwhile, NVIDIA's RTX Spark rollout shows the hardware side of this trend continuing to push AI compute outward from the data center toward the desktop.