Today’s episode starts with governments treating AI dependency as a policy object: search ranking rules, compute gaps, procurement choices, and summit politics all became part of the same operating problem for builders.

• The UK Competition and Markets Authority moved further toward conduct requirements for Google Search and AI Overviews, putting ranking and self-preferencing directly into the AI search conversation.
• The European Commission’s Digital Decade report framed Europe’s digital capacity gaps as an investment and dependency problem rather than only a competitiveness scorecard.
• The Guardian’s France reporting gave the policy story a concrete procurement example: moving away from Palantir toward domestic tooling.
• Fireworks AI, Parasail, and other providers turned GLM-5.2 into a day-zero availability story, with claims around long context and coding benchmarks that developers still need to test in their own loops.
• Qwen-RobotNav and MuseVLA show physical AI moving toward reconfigurable navigation, sensor selection, and embodied evaluation rather than demo-only robotics.
• The Rift deception paper offers a useful safety-research coda: internal signatures may help separate a model that is wrong from a model that is hiding what it knows, with caveats the authors state plainly.

Today’s episode starts with the reported SpaceX agreement to buy Anysphere, then follows the money, access rules, model releases, and capacity work around the tools developers now use every day.

• Reuters on SpaceX and Anysphere gives the day’s lead: a reported $60 billion stock deal for the company behind Cursor, with product consequences that remain unannounced.
• Techmeme’s OpenAI financials cluster surfaces reported audited figures for 2025 spending, including the research, development, and sales lines that make frontier AI a capital story.
• The Verge, Axios, and The Register add fresh reporting to the Anthropic Fable and Mythos access fight, including the disputed technical basis for the government intervention.
• NVIDIA’s Nemotron 3 Ultra paper, Open-SWE-Traces, CoAgent, and FragFuse make the research segment concrete: new models, coding traces, concurrency protocols, and memory-security attacks are all arriving at once.
• The GitHub capacity report and France’s Mistral-backed state-services plan show AI demand appearing as reliability work and procurement choices, not only model announcements.

Monday's Braid follows the same dependency from three angles: frontier model access is becoming political, agent reliability is moving into runtime controls, and policy is showing up as procurement rules and platform obligations.

• Axios and The Verge add reporting on Anthropic's Fable and Mythos shutdown, which keeps the weekend's model-access story focused on communication, procurement, and government pressure rather than another broad retelling.
• When Errors Become Narratives, the GNN tool-deference paper, and Minim turn the agent segment toward runtime behavior: plausible false outputs, indiscriminate tool trust, and privacy filtering before UI state leaves a device.
• The UK DSIT letter and Japan's Digital Agency guideline show AI policy arriving through regulator instructions and government purchasing practice.
• Techmeme's Enflame item, Rest of World's China open-source interview, and Two Minute Papers on Nvidia's open-weight model keep the open-model question grounded in chips, distribution, and local fallbacks.
• TechCrunch, Techmeme's FBI and Google item, and Al Jazeera round out the episode with labor pressure and concrete misuse stories that shouldn't be flattened into generic AI discourse.

Today’s episode follows the Anthropic Fable and Mythos cutoff from the first shock into the harder questions: who triggered the action, what the technical evidence actually proves, and what builders do when access to frontier models becomes a policy-dependent dependency.

• The Verge and Axios give the follow-up reporting on Amazon, the White House, and Anthropic, which moves the story from access shock into a dispute over evidence, trust, and export control.
• TechCrunch tracks the India reaction, where the cutoff becomes a national-dependence question rather than only a vendor incident.
• Techmeme’s labor roundup points to employers citing AI in about 88,000 U.S. job cuts through May, a number that is useful only if we separate layoff attribution from proven causality.
• Indian Express and TechCrunch cover KPMG pulling an AI report over apparent hallucinations and fake citations, a practical warning about professional evidence chains.
• Axios anchors the infrastructure section on power-market strain, while OpenRouter’s Fusion announcement and the weekend developer-tool links show how builders are adapting in smaller, more immediate ways.

Today's episode starts with Anthropic suspending access to Claude Fable 5 and Claude Mythos 5 after a reported U.S. government directive, then follows the practical consequence for builders: hosted model access is becoming part of compliance, infrastructure, legal discovery, and enterprise deployment design.

• Anthropic status incident anchors the lead: Fable 5 and Mythos 5 access was suspended, and Anthropic said it was working to restore access.
• Techmeme coverage of the export-control order gives the policy context around the reported government directive and the jailbreak evidence being cited.
• Open source AI must win captures the fast developer reaction: local and open models are being treated less like ideology and more like fallback architecture.
• The Guardian on the UK AI hardware push shows the other side of state involvement: governments are trying to fund chips, talent, and national capacity, not only restrict access.
• CNBC on state attorneys general and OpenAI brings the domestic legal track into view, where discovery can force operational claims into the record.
• Forbes on OpenAI and Ona points to the enterprise-agent deployment question: where the agent runs is now part of the product.
• ZDNET on OpenAI and Visa adds the payments version of the same story, where permissioning and reversibility matter as much as the model.

Today’s episode starts with Google’s WebMCP proposal, then follows the same question through open coding models, agent safety papers, China-facing hardware and robotics supply chains, AI mistakes in professional work, and ordinary developer security.

• Tara Agyemang’s AI Engineer talk on WebMCP gives the day its lead artifact: websites may need to expose actions directly to agents instead of making agents infer intent from pixels and DOMs.
• Moonshot AI’s Kimi K2.7-Code model page makes token efficiency part of the coding-model comparison, which matters when developers are paying for long agent runs.
• The agentic framework safety paper argues that common agent frameworks do not provide native structural containment guarantees, and its memory-poisoning experiment shows why framework behavior has to be tested separately from model behavior.
• The SMSR memory-poisoning paper proposes signed memory plus randomized retrieval as a more formal defense for persistent agent memory.
• Techmeme’s Nvidia-China item and its humanoid robot supply-chain item keep the infrastructure story grounded in chips, factories, and availability claims rather than model demos alone.
• Forbes’ court-sanctions story shows AI drafting running into a professional audit boundary, with lawyers removed after hallucinated legal citations appeared in filings.
• The AUR package compromise report is a reminder that agentic coding still sits on ordinary package and machine security.

Today's episode starts with Anthropic making a hidden Claude Fable 5 safeguard visible, then follows the same operational question into data centers, agents, search liability, robotics, and research systems: once AI becomes infrastructure, who can see the rule that changed the behavior?

• ClaudeDevs announced that flagged frontier-model-development requests will visibly fall back to Opus 4.8, turning an invisible safeguard into a user-facing signal.
• The Verge reported the apology and backlash around hidden Fable safeguards, which matters because researchers were evaluating behavior they could not clearly observe.
• Axios, The Guardian, and Al Jazeera show data-center politics moving from local siting disputes toward national policy over heat, power, water, and permitting.
• MIT Technology Review and same-day agent-governance papers point to a practical agent problem: identity, authority, refusal, and ownership after a system has access.
• Indian Express flags a court-risk signal around Google AI Overviews, where summary UI can turn into a liability surface.

Today's episode starts with the Trump administration reportedly telling CAISI to stop publishing public model assessments, then follows the same trust problem through compute deals, TCS's hiring plans, Anthropic's access terms, AWS Bedrock retention questions, and a small set of agent-security papers.

• Techmeme's CAISI roundup points to reporting that public model assessments have been halted while a new executive order is implemented; that changes the shared evidence builders can cite.
• Techmeme's OpenAI Ohio item and TechCrunch on Meta and Reliance show capacity turning into power contracts, geography, and financing.
• Techmeme's TCS item captures Tata chairman N. Chandrasekaran talking about agents as a hiring and workforce planning issue, not just a productivity demo.
• Anthropic's Mythos-class data-retention note makes the enterprise boundary more concrete: different clouds mean different retention and access paths.
• GitInject, the Interlocutor Effect paper, CIAware-Bench, and deployment-time memorization make the research tail practical: agents fail at the seams where code, memory, privacy, and oversight meet.

One morning's arXiv listing dropped close to twenty agent papers, and almost none of them are about making agents more capable. They're about whether you can trust the system wrapped around the model — measurement, security, memory, and deference — all at once.

• Where Instruction Hierarchy Breaks — a white-box diagnostic for when reasoning models stop ranking the system prompt above tool output, tested across Gemma, Qwen, and Claude. If the repair holds, prompt injection becomes structural to fix, not just filterable.
• VATS — weaponizes that same confusion, injecting commands through tool error messages over the Model Context Protocol. The error path is the door most teams never locked.
• Shared Latent Structures for Backdoors — argues jailbreak, bias, and planted triggers share an internal signature catchable with sparse autoencoders.
• Beyond Goodhart's Law (MAC-Bench), Online Agent-as-a-Judge, and PACE — three attempts to keep evaluation honest when the thing you're testing can learn the test.
• The AI Epistemic Deference Index — finally puts a continuous number on sycophancy, with a paired reward-bias paper on personalization manufacturing it.
• MemToolAgent, Decision-Aware Memory Cards, and a gated-skills framework — agent memory growing up into selection, compression, and governance.
• Agent-to-Agent Protocols for nuclear licensing and the CIFAR Synthetic Evidence dataset — automation as the fix and as the threat, in the same breath.
• Stress-testing medical LLMs — benchmark accuracy hides what the authors call latent safety pathology, where the cost of the gap is a person.

The consequential part of an AI system keeps moving out of the model and into the wrapper around it — the cooling loop, the org chart, the config file, the ownership structure — and the tools we use to trust that wrapper are running behind it. Five stories, one recurring tension.

• The Guardian finds about two-thirds of 809 planned US datacenters are slated for drought-hit land; closed-loop cooling saves water but trades it for fossil power that needs water of its own.
• OpenAI's enterprise talks feature banks rebuilding their orgs: Allica Bank collapsing roles into "squadlets," Erste Group budgeting for a full platform rewrite every 18 months, plus ChatGPT-in-Excel Skills and Codex — held against one engineer's MCP catalog server.
• The Miasma worm: one dropper wired into seven config files across Claude Code, Gemini, Cursor, VS Code, npm, Composer, and Bundler — opening a cloned repo becomes an execution event.
• Schneier and Nathan Sanders argue against Bernie Sanders' equity-stake plan, proposing energy taxes and an AI Public Option instead — set against Korea's GPU program and NVIDIA's UK sovereign-AI post.
• Two arXiv papers on measuring safety too late: Attack Selection shows strategic timing drops measured control safety 20-28 points, and Don't Just Fix It in Post argues the science belongs in training dynamics, not the finished snapshot.