#6 — Anthropic Mythos 5, Fable 5, Canada $2.3B AI strategy

What happens when you relax an AI's safety guardrails to help it generate scientific hypotheses? That's one of the stories on Research Agent this week. We'll start with the new models from Anthropic. [LEAD] Anthropic released Claude Mythos 5 Anthropic introduced Claude Mythos 5, a large language model with relaxed guardrails. This model is accessible to a limited number of organizations. It has demonstrated capability in generating novel scientific hypotheses. SOURCE MATERIAL: (Anthropic sets AI performance records with new Mythos 5, Fable 5 ..., 2026-06-09): Anthropic PBC has introduced Claude Mythos 5 and Claude Fable 5, two new large language models that reportedly outperform competitors across various benchmarks. Mythos 5, with more relaxed guardrails, is available to a limited number of organizations and has shown the ability to produce novel scientific hypotheses. Fable 5, broadly available, includes stricter guardrails and has set a record on the SWE-Bench Pro programming benchmark. [LEAD] Anthropic released Claude Fable 5 Anthropic introduced Claude Fable 5, another large language model featuring stricter guardrails. This model is broadly available and established a record on the SWE-Bench Pro programming benchmark. SOURCE MATERIAL: (Anthropic sets AI performance records with new Mythos 5, Fable 5 ..., 2026-06-09): Anthropic PBC has introduced Claude Mythos 5 and Claude Fable 5, two new large language models that reportedly outperform competitors across various benchmarks. Mythos 5, with more relaxed guardrails, is available to a limited number of organizations and has shown the ability to produce novel scientific hypotheses. Fable 5, broadly available, includes stricter guardrails and has set a record on the SWE-Bench Pro programming benchmark. [LEAD] Canadian federal government allocated $2.3 billion for AI for All strategy. The Canadian federal government launched the AI for All strategy. This initiative includes a $2.3 billion funding allocation. The strategy aims to support various AI-related activities across Canada. SOURCE MATERIAL: (Amii gets share of federal funding in AI strategy - Taproot Edmonton, 2026-06-09): The federal government's new AI for All strategy allocates $2.3 billion, part of which the Alberta Machine Intelligence Institute (Amii) will share with sister institutes. This funding supports research positions, an AI Safety Institute, and the commercialization of AI research. Amii CEO Cam Linke emphasized the strategy's importance in closing Canada's AI deployment gap through literacy and training programs. [LEAD (FOLLOW-UP)] Loop engineering replaces manual prompting with goal-based automation in AI systems Loop engineering replaces manual prompting in AI systems with automated, goal-driven processes. This method involves agents executing actions, observing outcomes, determining next steps, and iterating until a set objective is met. It is essential for AI coding agents to handle complex multi-step tasks, integrate real-world feedback, and perform iterative refinements. Key elements for successful loop engineering include defining clear goals, providing appropriate tools, managing context, establishing termination criteria, and implementing error handling. SOURCE MATERIAL: (What Is Loop Engineering? The New Meta for AI Coding Agents ..., 2026-06-09): Loop engineering replaces manual prompting with goal-based automation in AI systems, allowing agents to act, observe results, decide next steps, and repeat until a goal is met. This iterative approach is crucial for AI coding agents to handle multi-step tasks, real-world feedback, and iterative refinement, as opposed to single-shot prompting. A well-engineered loop requires a clear goal, a relevant toolset, context management, termination logic, and robust error handling. [LEAD] A survey proposes five evaluation dimensions for AutoResearch autonomy. The survey suggests five key dimensions for evaluating AutoResearch autonomy: novelty, validity, impact, reliability, and provenance. These dimensions provide a structured approach to assessing the effectiveness and credibility of AI-powered scientific workflow automation. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [LEAD] AMAR automates academic research workflows with specialized AI agents The AMAR system, an end-to-end web-based platform, automates the complete academic research workflow. It integrates seven specialized AI agents for tasks ranging from literature discovery and experiment execution to draft generation. This system aims to enhance research reproducibility and rigor by decomposing complex tasks across its agent pipeline. SOURCE MATERIAL: (AMAR: An Autonomous Multi-Agent Researcher for End to End Automated Scientific Literature Review and Draft Generation, 2026-05-21): We present AMAR (Autonomous Multi-Agent Researcher), an end-to-end web-based system that automates the complete academic research workflow — from literature discovery through experiment execution to draft generation. AMAR orchestrates a pipeline of seven specialized AI agents: a Searcher Agent that retrieves papers from arXiv, OpenAlex, CrossRef, and IEEE Xplore APIs; a Summarizer Agent; a Critic Agent that identifies research gaps; a Developer Agent that synthesizes executable Python experiment code; an Experimenter Agent that executes generated code in an isolated Docker sandbox; a Verifier Agent that validates citations and results against academic databases; and a Writer Agent powered by [LEAD] StructAgent automates cryo-EM model building StructAgent is a user-guided multi-agent system designed for cryo-electron microscopy (cryo-EM) model building and refinement. It integrates a domain agent for literature-grounded structural reasoning with an execution agent that operates local software. The system tracks state, manages failures, and records provenance, with expert approval gates for major model changes. SOURCE MATERIAL: (StructAgent: Orchestrating Cryo-EM Model Building and Refinement with a Multi-Agent LLM System, 2026-05-18): Abstract Building and refining cryo-EM atomic models often requires long, project-specific workflows that combine map inspection, prior structural knowledge, restraints, refinement, validation and expert review. Existing programs perform many individual operations, but coordinating them across iterative model-building sessions remains manual and difficult to audit. We present StructAgent, a user-guided multi-agent resource for cryo-EM model building and refinement. StructAgent couples a domain agent for literature-grounded structural reasoning with an execution agent that runs local software, tracks state, recovers from failures and records provenance. Expert approval gates control major mod [NOTABLE (FOLLOW-UP)] AI models accelerate scientific discovery but may overlook novel findings Research indicates that AI, particularly through transfer learning, can expedite the search for new physical laws, reducing the need for extensive simulations. However, this acceleration comes with a drawback: AI's reliance on familiar data can cause it to miss genuinely new discoveries. This behavior stems from the models' confidence in established patterns, which can prevent the recognition of unexpected results. SOURCE MATERIAL: (AI could uncover new physics faster but there's a surprising catch ..., 2026-06-11): AI may help uncover new laws of the universe faster—but sometimes it's so confident in what it already knows that it misses the surprise. Scientists found that transfer learning can make the search for new physics in the universe much faster, slashing the need for expensive simulations. Yet the approach can backfire when AI relies too heavily on familiar patterns, potentially missing evidence of something truly new. [NOTABLE] Relai Inc. launched its verifiable continual learning platform. Relai Inc. introduced a verifiable continual learning platform. This platform converts AI agent failures and feedback into effective learning environments. Its goal is to improve knowledge and correct mistakes, addressing unpredictable failures and regressions by validating improvements and routing fixes to the appropriate agent layer. SOURCE MATERIAL: (Exclusive: Relai raises $6.9M to enable verifiable and continuous learning for AI agents, 2026-06-10): Artificial intelligence infrastructure startup Relai Inc. has secured $6.9 million in funding to enhance the reliability of autonomous AI agents for enterprises. The company also launched its verifiable "continual learning" platform, which aims to transform AI agent failures and feedback into reliable learning environments to improve knowledge and rectify mistakes. Relai addresses the challenge of unpredictable AI agent failures and silent regressions by validating improvements against prior environments and routing fixes to the appropriate layer of the agent's stack. [NOTABLE] AI for All strategy supports AI research positions, an AI Safety Institute, and commercialization. The AI for All strategy details specific areas of support. It focuses on creating new AI research positions. The strategy also includes the establishment of an AI Safety Institute. Furthermore, it aims to commercialize AI research findings. SOURCE MATERIAL: (Amii gets share of federal funding in AI strategy - Taproot Edmonton, 2026-06-09): The federal government's new AI for All strategy allocates $2.3 billion, part of which the Alberta Machine Intelligence Institute (Amii) will share with sister institutes. This funding supports research positions, an AI Safety Institute, and the commercialization of AI research. Amii CEO Cam Linke emphasized the strategy's importance in closing Canada's AI deployment gap through literacy and training programs. [NOTABLE] President Trump issued an Executive Order on AI cybersecurity and NSPM-11. President Trump issued an Executive Order (EO) on AI cybersecurity and NSPM-11, indicating a preference for voluntary industry engagement. These directives aim to accelerate AI adoption across the national security enterprise. The EO reduces the government's pre-release access period for frontier AI models from 90 to 30 days and explicitly rejects mandatory licensing. SOURCE MATERIAL: (Trump Administration and House Lawmakers Launch New AI Governance Initiatives, 2026-06-11): On June 2, 2026, President Trump issued an EO on AI cybersecurity and, days later, NSPM-11, signaling the administration's preference for voluntary engagement with industry partners while accelerating AI adoption across the national security enterprise. The EO reflects a lighter-touch approach than earlier drafts, shortening the government's pre-release access period for certain frontier AI models from 90 days to 30 days and expressly rejecting mandatory licensing or preclearance requirements. On June 4, 2026, Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) released the Great American AI Act discussion draft, the most comprehensive bipartisan AI framework proposed to date, showing their in [NOTABLE (FOLLOW-UP)] Current AI systems in scientific workflows exhibit fragmentation and struggles. Existing AI systems for scientific workflows are fragmented, varying in autonomy, domain scope, and execution environments. These systems still struggle with essential aspects such as evidence preservation, reproducibility, provenance tracking, and robust cross-domain application. They also face limitations in accountability and scientific closure. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] A survey defines "AutoResearch" as the spectrum of AI-powered scientific workflow automation. A survey introduces the concept of "AutoResearch," defining it as the entire developmental spectrum of AI-powered automation within scientific workflows. This framework helps categorize the varying degrees of AI involvement in research processes. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] A survey defines "Vibe Research" as human-steered prompt-based assistance within scientific workflows. The survey identifies "Vibe Research" as a specific region within AutoResearch. This region is characterized by human-steered, prompt-based assistance and requires human verification of execution. It contrasts with more autonomous AI-led systems. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] A survey finds AutoResearch autonomy is domain-conditioned. The survey concludes that the credibility of AutoResearch autonomy depends on the specific scientific domain. Autonomy is more reliable in structured, executable, and rapidly verifiable settings. Conversely, its application is limited in contexts that are embodied, delayed, heterogeneous, ethical, or institutionally accountable. SOURCE MATERIAL: (AutoResearch AI: Towards AI-Powered Research Automation for Scientific Discovery, 2026-05-22): Scientific research is being reshaped by AI systems that move beyond isolated assistance toward longer-horizon workflows spanning literature grounding, hypothesis generation, experimentation, validation, reporting, and revision. This shift marks a transition from task-level AI for science to workflow-level research automation. Yet current systems remain fragmented, differing in autonomy, domain scope, execution environment, validation mechanism, and human oversight, while still struggling with evidence preservation, reproducibility, weak-direction rejection, provenance tracking, cross-domain robustness, and accountable scientific closure. This survey examines these developments through AutoR [NOTABLE] StructAgent demonstrates capabilities in cryo-EM model refinement StructAgent's capabilities were demonstrated across three case studies. It refitted a 64-chain proteasome from an earlier template. The system also audited 530 ribosomal metal-ion sites. Additionally, StructAgent guided a chemically ambiguous ligand fit in a folate-metabolism enzyme. SOURCE MATERIAL: (StructAgent: Orchestrating Cryo-EM Model Building and Refinement with a Multi-Agent LLM System, 2026-05-18): Abstract Building and refining cryo-EM atomic models often requires long, project-specific workflows that combine map inspection, prior structural knowledge, restraints, refinement, validation and expert review. Existing programs perform many individual operations, but coordinating them across iterative model-building sessions remains manual and difficult to audit. We present StructAgent, a user-guided multi-agent resource for cryo-EM model building and refinement. StructAgent couples a domain agent for literature-grounded structural reasoning with an execution agent that runs local software, tracks state, recovers from failures and records provenance. Expert approval gates control major mod That's it for today. We'll have more on AI systems for research next time. From Research Agent, thanks for listening.