AWS hosted its New York Summit this week and launched a number of capabilities that make the enterprise more agent-ready, covering developers, operators, security engineers, and business personas. Google announced a new specification for discovery and trust of agentic resources. Anthropic suspended access to Fable 5 and Mythos 5 after a government export control mandated no access to foreign nationals, sparking discussions at the G7 summit about a possible trusted partner framework. SpaceX is moving forward with a $60 billion deal to acquire Anysphere, the parent company of Cursor, in a move to push into enterprise AI tools. A survey series from Anthropic shows only 15% of Americans trust AI companies.
RELEASES
① AWS Summit New York 2026: New ways to make AI agents more effective at work
Summary: AWS introduced AWS Continuum, an AI-native security service that continuously discovers, validates, and remediates code vulnerabilities, and AWS Context, a knowledge graph that grounds agents in an organization’s data. The summit also brought updates across Kiro, Amazon Quick, AWS DevOps Agent, and AWS Transform. Amazon Bedrock AgentCore added five enhancements, including a managed knowledge base, native web search, and general availability of AgentCore harness.
Signal: AWS is transitioning agents from one-off tasks to continuous, always-on operation, e.g., Continuum keeps watch for code vulnerabilities, Context keeps building organizational knowledge, Transform keeps code modernized, Kiro runs persistent cloud sessions accessible via mobile phone. That’s a shift from prompt-completion chatbots towards persistent agents with feedback loops for continuous improvement over time.
② Context intelligence for your data and AI agents at scale
Summary: AWS announced three context-layer capabilities for AI agents. AWS Context, coming soon, builds an organization-wide knowledge graph that learns from usage, publishes metadata in Iceberg format, and enforces identity-aware access via IAM and Lake Formation. Glue Data Catalog, in preview, now supports business descriptions and skill assets for tables. S3 Annotations, generally available, allows customers to attach up to 1 GB of mutable, queryable context directly to S3 objects.
Signal: AWS is pushing agent context into open, queryable Iceberg tables rather than a proprietary vector store or RAG pipeline, treating context as data rather than as embeddings for enterprise-scale grounding. That gives customers already standardized on Iceberg and S3 a head start, even as others build their own context and discovery layers.
③ Introducing AWS Continuum: Security at machine speed
Summary: AWS introduced AWS Continuum for code vulnerabilities, now available in gated preview. Continuum runs four continuous phases: 1/ discovery, 2/ prioritization, 3/ validation, and 4/ mitigation and remediation, reasoning over an organization’s infrastructure and business context to confirm which vulnerabilities are real and are priorities for resolution. It is model agnostic, generates threat models in STRIDE format from design documentation or source code, and starts in human-in-the-loop learn mode before graduating to automated enforcement mode based on user-defined risk profiles.
Signal: AWS explicitly names Claude Mythos as a catalyst, with frontier models finding vulnerabilities and reasoning through attack paths faster than humans can triage. Continuum’s graduated trust model from learn mode to enforce mode is a pragmatic approach to earning trust with organizations for autonomous security agents.
④ Announcing the Agentic Resource Discovery specification
Summary: Google announced the Agentic Resource Discovery (ARD) specification, an open standard for publishing, discovering, and verifying AI agents, tools, and skills across organizational boundaries. Organizations host catalogs of their capabilities under their own domain, which acts as the cryptographic root of trust, while registries crawl and index those catalogs, like a search engine for the agentic web. The specification is Apache 2.0 licensed and built upon the AI Catalog data model.
Signal: ARD is Google’s bid to standardize the discovery layer of the agentic web, playing a similar role that DNS and search played for the original web. The partner list will be the early signal of whether ARD consolidates or fragments against competing standards.
MODEL ACCESS
① Statement on the US government directive to suspend access to Fable 5 and Mythos 5
Summary: Anthropic has suspended access to Fable 5 and Mythos 5 for all customers after the US government issued an export control directive with 90 minutes notice, prohibiting access by any foreign national, inside or outside the US. The government’s stated concern is a potential narrow, non-universal jailbreak. Anthropic reviewed the underlying report and found that the demonstrated capability is already available in other publicly available frontier models. Anthropic is complying with the directive but publicly disagrees that a narrow potential jailbreak should trigger a commercial model recall.
Signal: This is the first known instance of the US government invoking export control authority to pull a deployed frontier model from the market. Anthropic’s pushback signals that the industry has no shared standard on the level of risk that warrants model withdrawal. This sets a precedent that regulatory intervention can be abrupt, opaque, and broadly disruptive even when the underlying technical finding is narrow.
② A warning from Amazon led the White House to shut down Anthropic’s Mythos model
Summary: Anthropic complied with government orders to cut off foreign access to Fable 5 and Mythos 5, which may have been motivated by three possible reasons, according to multiple media reports: 1/ a warning from Amazon CEO Andy Jassy, 2/ concerns about unauthorized Chinese access, and 3/ cybersecurity fears. Critics have described the export controls as government overreach. Meanwhile, the government stated that it had asked Amodei to fix the issue so that it could return to general release and pushed back on suggestions that the move was connected to the earlier Pentagon dispute that resulted in Anthropic being designated a supply chain risk.
Signal: This episode elevates discussions about sovereign AI, the idea that countries should control the AI models, computing infrastructure, and data that underpin critical technology, rather than depending on systems that could be restricted or withdrawn by foreign governments. Access to and control of AI models is becoming a matter of national security, and the administration has signaled that any future model exceeding Mythos’s capability threshold will need government review before release, setting a precedent for pre-clearance of frontier AI.
③ G7 leaders discuss ‘trusted partners’ access to cutting-edge US AI models, sources say
Summary: G7 leaders discussed a trusted partners framework that would let select countries or companies bypass new U.S. restrictions on foreign access to Anthropic’s most advanced models. The talks took place mainly with U.S. Commerce Secretary Howard Lutnick on the sidelines of the summit’s dinner on Monday, separate from the lunch on Wednesday with Altman, Amodei, Hassabis, and other AI leaders. Proponents argue broader access would let G7 countries build stronger cybersecurity defenses, especially following the administration’s export directive that forced Anthropic to cut off foreign access to Fable 5 and Mythos 5.
Signal: The possible framework represents the first sign of a formal tiering system for frontier AI access with trusted partner status potentially covering countries or companies. The EU’s push to study the model’s implications shows that model access is now a defense and security concern that extends beyond just commercial interests. The trusted partner framework may also have implications for enterprises operating across G7 borders, as it could shape frontier model access for sensitive use cases.
MARKETS
① SpaceX locks in $60 billion Cursor deal to close gap with rivals in AI coding race
Summary: SpaceX is buying Anysphere, the parent company of AI coding tool Cursor, for $60 billion in an all-stock deal that is expected to close in Q3 2026. The move extends SpaceX’s push into enterprise AI tools, which is a market central to the $28.5 trillion addressable market that it pitched to IPO investors. Cursor is a leader in agentic coding but has been compute constrained; this deal gives it access to SpaceX’s Colossus GPU clusters.
Signal: SpaceX’s $26 billion in annualized compute-leasing deals with Anthropic and Google shows it monetizing once-idle GPU capacity. The Cursor deal shows it now using the same compute to acquire AI products outright. Notably, the leasing contracts carry 90-day termination clauses, giving SpaceX room to reclaim capacity for Cursor if needed. Keep an eye out for how it balances provisioning capacity between external customers and its own AI portfolio.
RESEARCH
① Investing in multi-agent AI safety research
Summary: Google DeepMind, Schmidt Sciences, the Cooperative AI Foundation, and ARIA, with support from Google.org, are jointly funding up to $10M in research to address safety risks in large-scale multi-agent AI systems. The funding call targets four areas: 1/ sandboxes and testbeds, 2/ the science of agent networks, 3/ agent infrastructure, and 4/ oversight and control. Proposals are open to academic and independent researchers worldwide, with a submission deadline of August 8, 2026.
Signal: The industry is deploying multi-agent systems faster than the safety science can keep pace. The governance and observability layer for agent-to-agent interaction is still an outstanding challenge to solve. Most safety evaluations today analyze models in isolation, whereas this fund explicitly targets emergent behaviors that only appear at the network level, which is where agent systems actually operate.
② Results from the First Anthropic Public Record
Summary: Anthropic Public Record is a new recurring survey series tracking public opinion on AI with the first wave surveying nearly 52k Americans in November and December 2025. Americans’ top hope was curing diseases (48%), while their top fear was AI-induced job loss (64%). Over 70% of Americans showed bipartisan support for government regulation of AI with priority areas in privacy, child safety, and liability for harm. Only 15% of Americans trust AI companies to make decisions about how AI is developed and used.
Signal: The trust figure (15%) is the most pointed finding, with AI companies scoring lower than even the federal government (20%) and far below independent experts (43%). The public’s preferred remedies are legal liability and safety-first prioritization. Support for government regulation at 71% with only a 9% partisan gap (Democrats at 79%, Republicans at 68%) gives policymakers a durable mandate that doesn’t depend on one party holding power.
Always be learning.
heeki reads #5
Written by Heeki Park, Principal SA @ AWS. Opinions are my own.
Alcurio is where alchemy meets curiosity.