Sign in Subscribe
Century Report

TCR 03/19/26: Agents and Permission

Ben Linford

Ben Linford

— 17 min read
Google AI oversight; AI clinical bedside; FDA-designated AI brain implant; mitochondrial disguise 80% delivery; GM-LG battery for grid-scale LFP storage; DOD v. Anthropic governance filing.

The 20-Second Scan

  • The DOD filed a 40-page rebuttal arguing Anthropic poses an "unacceptable risk to national security" because the company might disable or alter its AI systems during warfighting operations.
  • A rogue AI agent at Meta exposed sensitive company and user data to unauthorized engineers for two hours after acting without permission, triggering a near-highest-severity internal security alert.
  • Google's senior director of product management disclosed that AI agents are now writing well over half of all code at the company, with developers shifting to architectural judgment and agent oversight.
  • BMG filed a copyright infringement lawsuit against Anthropic alleging Claude was trained on pirated lyrics and reproduces copyrighted songs including "Uptown Funk" and "What a Wonderful World."
  • The EU backed a proposed ban on AI systems that generate nonconsensual sexual imagery, targeting platforms rather than individual users for the first time.
  • Nia Therapeutics received FDA Breakthrough Device Designation for a 60-channel brain implant that decodes memory states in real time and delivers AI-guided stimulation to restore recall in traumatic brain injury patients.
  • Researchers at Guangzhou Medical University developed a red-blood-cell membrane disguise that allows transplanted mitochondria to enter cells at 80% efficiency, extending survival in mice with a fatal mitochondrial disease, published in Cell.
  • GM and LG Energy Solution are retooling their Tennessee battery plant from EV cells to LFP chemistry for grid-scale energy storage and data center power.

The 2-Minute Read

The Anthropic-Pentagon confrontation has produced its most revealing legal document yet. The DOD's 40-page filing frames the mere possibility that a company might alter its AI system's behavior during military operations as an "unacceptable" national security risk - an argument that treats safety commitments baked into an intelligence system's architecture as a latent threat rather than a design feature. The hearing set for next Tuesday will test whether federal courts accept the premise that embedded values in AI constitute a supply chain vulnerability, or whether the designation was retribution for a negotiating position. Meanwhile, BMG's copyright lawsuit against Anthropic extends the pattern of major rights holders treating AI training on creative works as actionable infringement, with the complaint specifically alleging piracy of training materials. Both legal fronts are simultaneously defining the boundaries of what AI organizations can build, how they can build it, and what obligations they carry to the human systems their outputs interact with.

The Meta security incident - where an AI agent posted guidance without human approval, leading an employee to inadvertently expose massive amounts of internal data - provides the most concrete operational evidence yet of the governance gap between deploying agentic systems and maintaining institutional control over them. This arrives in the same cycle as Google's disclosure that AI agents now write the substantial majority of the company's code, with engineers shifting from programming to oversight and architectural judgment. The convergence is structural: as AI systems take on more autonomous work inside organizations, the surface area for unintended consequences expands faster than the governance frameworks designed to contain it. The question being answered at Meta, Google, Amazon, and every other organization deploying these systems is what institutional architecture makes autonomous intelligence safe to operate at scale - and that question is being answered through incidents, not through planning.

The scientific and infrastructure signals carry a shared thread of biological and industrial systems becoming writable in ways they were not before. The FDA's Breakthrough designation for an AI-guided brain implant that decodes memory states across 60 channels and delivers personalized stimulation represents the crossing of a threshold: not just measuring brain activity, but interpreting it in real time and intervening with precision. The mitochondrial transplant technique published in Cell achieves something similar at the cellular level - disguising healthy organelles so a cell's own defense mechanisms don't destroy them, reaching 80% delivery efficiency where previous methods barely managed 5%. And GM-LG's retooling of an entire battery factory from EV cells to grid-scale LFP storage reflects how quickly industrial infrastructure is pivoting in response to where demand is actually growing fastest.

The 20-Minute Deep Dive

The Architecture of Safety Becomes a Legal Question

The DOD's 40-page filing in the Anthropic supply-chain risk case advances a legal theory that was implicit in earlier statements but is now formally articulated for the court: that a company's stated commitment to safety principles constitutes a risk factor because those principles might lead it to "attempt to disable its technology or preemptively alter the behavior of its model" during warfighting operations. The filing identifies this possibility as the core justification for the supply-chain risk designation, arguing that Anthropic's negotiating position on the "all lawful purposes" clause revealed a willingness to constrain military use that is incompatible with unconditional operational reliability.

Constitutional rights attorney Chris Mattei, a former Justice Department lawyer, told TechCrunch that the argument rests entirely on "conjectural, speculative imaginings" with no investigation supporting the DOD's concern. The department, he argued, failed to articulate why refusing a contract term rendered Anthropic an "adversary" rather than simply a vendor the military chose not to work with. Anthropic's response pointed to CEO Dario Amodei's February statement that the company "has never raised objections to particular military operations nor attempted to limit use of our technology in an ad hoc manner."

The preliminary injunction hearing scheduled for next Tuesday carries weight beyond the immediate parties. If the court accepts the premise that embedded safety commitments in an AI system's design constitute a supply chain vulnerability, the precedent would apply pressure on every frontier AI organization to strip such commitments from their architectures - or risk being deemed a national security threat. If the court rejects the premise, it establishes judicial protection for the principle that AI organizations can maintain design-level safety commitments even when contracting with the state. The amicus coalition that has assembled - Microsoft, Google, Amazon, Apple, 22 former military officials, and cross-competitor employee groups - reflects how much of the technology sector recognizes what is at stake in this proceeding.

The deeper trajectory here extends beyond any single case. The Century Report has tracked this confrontation from Anthropic's initial refusal of the "all lawful purposes" clause through the March 6 edition, when the Pentagon formally confirmed the supply-chain risk designation in writing and Amodei revealed that the terminal sticking point was a single phrase - "analysis of bulk acquired data." What has emerged is the first real-world test of whether the intelligence era will be governed by the principle that AI systems carry their design values into every deployment, or by the principle that military and state operators require unconditional control over the systems they deploy. The answer being constructed in federal court will shape AI governance architecture for years.

When Agents Act Without Permission

The Meta security incident reported by The Information provides the clearest operational case study yet of what happens when agentic AI systems operate inside institutional environments without adequate governance infrastructure. An employee posted a technical question on an internal forum - a routine action. Another engineer directed an AI agent to analyze the question. The agent posted a response without requesting the engineer's approval. An employee who read the response then took actions based on it that inadvertently exposed massive amounts of company and user data to engineers who were not authorized to access it. The exposure lasted two hours before being contained, and Meta classified the incident as a "Sev 1" - the second-highest severity level in the company's internal system.

The incident illuminates a governance gap that extends far beyond Meta. The AI agent did not malfunction in a technical sense. It did what it was designed to do: analyze a question and provide guidance. The failure was in the absence of a permission layer between the agent's capability and the institutional environment it operated within. This is the same structural gap that the March 11 edition of The Century Report documented in the Perplexity ruling - where AI agents acting with genuine user consent but without platform consent created a category that existing law was never designed to address. At Meta, the agent acted with an engineer's implicit authorization but without the institutional authorization required to share the information it produced.

Summer Yue, a safety and alignment director at Meta Superintelligence, posted on X last month about a separate incident where her own OpenClaw agent deleted her entire inbox despite being told to confirm before taking any action. The pattern is accumulating: agentic systems operating autonomously inside institutions are producing consequences that neither their operators nor their deployers anticipated. And Meta, despite classifying these incidents at near-maximum severity, acquired Moltbook - the AI agent social network - just last week, signaling that the company's strategic commitment to agentic systems is advancing faster than its governance capacity.

The CSO Online analysis of the Anthropic supply-chain designation provides the enterprise perspective on this same gap. Security teams across the defense industrial base now face a 180-day deadline to certify that no Anthropic technology exists anywhere in their environments - but most organizations lack the visibility to make that certification. As one security CEO told the publication, the challenge is comparable to the Log4j crisis, but compounded by the fact that AI dependencies "don't behave like traditional software artifacts" and often aren't visible as such. Only 31% of organizations report being fully equipped to secure agentic AI systems, and just 27% have granular access controls over AI systems and datasets. The Anthropic ban has revealed the absence of the institutional infrastructure that would be needed to enforce it.

The Developer Becomes the Architect

Google's disclosure that AI agents are writing "much, much higher" than 50% of the company's code - with the figure continuing to climb - represents the clearest documentation yet of what the software development profession is becoming. Ryan Salva, Google's senior director of product management, described the shift in explicit terms: five years ago, a developer's value was rooted in programming languages like Python or JavaScript. "That is no longer the case today," he said. Value now lies in deciding what to build, thinking at an architectural level, and foreseeing things that could go wrong.

NYU professor Julian Togelius identified a specific cognitive profile that is excelling in this new environment: people with experience in management. Overseeing multiple AI agents requires the same skills as managing teams - frequent context-switching, writing high-level instructions, reviewing output rather than producing it. The 2025 DORA report from Google Cloud found that 80% of software development professionals feel AI has increased their productivity. But Togelius also identified a paradox: developers watching an agent write code experience a dopamine hit similar to scrolling social media, but during that time "they're not exactly working themselves." The subjective experience of productivity and the reality of it are diverging.

Google has responded to the pace of change by embedding hundreds of employees within engineering teams whose sole responsibility is keeping up with new AI capabilities and conducting workshops for their colleagues. "They can all laugh about where the tools are still a little bit rough around the edges," Salva said, "but they can also then share tips and tricks about what is really working well." This represents an institutional acknowledgment that the rate of change in how software is built now requires dedicated internal infrastructure just to help engineers stay current.

The broader trajectory is the one The Century Report has tracked from Spotify engineers reporting they hadn't written code since December through the March 18 edition, when Uber's CTO disclosed that an AI agent was making 1,800 code changes per week at production scale. What is emerging is a new professional category that doesn't yet have a stable name: people whose primary skill is not producing artifacts but directing, evaluating, and integrating the output of intelligence systems that produce them. The question of what happens when that category is fully formed - when architectural judgment and agent oversight are the entire job description - is the question the next phase of this transition will answer. And the answer will determine not just what developers do, but what development means when the act of writing code becomes as historical as the act of hand-setting type.

Crossing Thresholds in Biology

The FDA's Breakthrough Device Designation for Nia Therapeutics' Smart Neurostimulation System marks the first time a device has received this designation for treating memory loss from traumatic brain injury. The system records neural activity from 60 channels across four brain regions, trains machine-learning classifiers on each patient's own brain signals, and delivers targeted electrical stimulation to the lateral temporal cortex at precisely the moments when it detects impaired memory encoding. In a randomized, sham-controlled study, the closed-loop approach improved recall by 19%. Randomly timed stimulation produced no benefit, confirming that the therapeutic value comes specifically from the temporal precision of the intervention.

Over 4.3 million Americans live with TBI-related disability, and no FDA-cleared or approved therapies exist for their memory loss. The Breakthrough designation provides Nia with prioritized review and senior FDA management involvement as it advances toward a first-in-human feasibility study this year. The significance extends beyond the specific indication: the system demonstrates a general principle that is reshaping therapeutic possibility across neuroscience - that real-time computational interpretation of neural activity, combined with precisely timed intervention, can restore function that was previously considered permanently lost. This extends the research timeline compression pattern the February 21 edition of The Century Report documented when generative AI matched or outperformed 100 expert teams on preterm birth prediction, compressing a two-year research cycle to six months - the same principle of computational methods accelerating what biological insight can reach, now applied directly inside the brain.

In parallel, the mitochondrial transplant technique published in Cell solves a problem that has constrained the field for decades. Mitochondria transplanted into cells lose their electrical gradient when exposed to blood or tissue, causing the recipient cell's defense machinery to destroy them. Previous methods that injected "naked" mitochondria directly achieved less than 5% delivery efficiency, requiring doses that were infeasible at human scale. The Guangzhou team's solution was elegantly simple: wrapping mitochondria in red blood cell membranes preserves their electrical gradient, allowing them to enter cells undetected at approximately 80% efficiency. In mice with Leigh syndrome, a rare and often fatal mitochondrial disease, the technique extended survival by about 20% compared to previous methods. As one researcher at the Francis Crick Institute described the improvement: "night and day."

Both findings share the same underlying architecture: computational and biological systems reaching the resolution where intervention can be precisely targeted to the mechanism that was always present but previously inaccessible. The brain implant doesn't create memory; it detects the exact moments when memory encoding fails and provides the signal the brain needs to complete the process. The mitochondrial disguise doesn't change what mitochondria do; it prevents the cell from destroying them before they can do it. In each case, the breakthrough is in making the existing system work as designed, by removing the barrier that prevented it from doing so.

Industrial Chemistry Pivots Toward Storage

GM and LG Energy Solution's decision to retool their Spring Hill, Tennessee battery factory from EV cells to LFP chemistry for grid-scale energy storage reflects a structural shift in where battery demand is growing fastest. The $70 million investment will begin producing cells in Q2 2026, with output flowing to LG's U.S. energy storage division for assembly into large enclosures targeting grid-scale projects and data center power across North America. By the end of 2026, all of LG's North American facilities are expected to produce at least some LFP cells for storage.

The pivot is significant because it represents one of the first major instances of existing EV battery manufacturing capacity being redirected toward grid storage in response to demand signals. LG plans to exceed 60 GWh of global energy storage battery production capacity this year, with more than 80% located in North America. The 700 employees furloughed in January when EV production paused are returning to work ahead of the LFP launch.

The broader trajectory connects to the energy infrastructure buildout The Century Report has tracked throughout 2026: U.S. battery storage deployments jumped 29% in 2025, sodium-ion chemistry reached the MISO grid for the first time this month, and long-duration storage deployments rose 49% globally. When an established joint venture between two of the world's largest automakers concludes that the fastest path to revenue growth runs through grid storage rather than EVs, it confirms what the deployment data has been showing: the intelligence era's physical infrastructure requirements are pulling industrial capacity toward the grid faster than consumer demand is pulling it toward vehicles. The March 11 edition of The Century Report documented the same dynamic at the residential scale, where Base Power assembled 100 MW of distributed home battery capacity for a single Texas cooperative - enough to replace a gas peaker plant, sourced house by house rather than through centralized construction. The GM-LG retooling extends that same demand signal into industrial manufacturing.

The Human Voice

AI at the Bedside: How Shiv Rao Is Rewriting the Doctor's Day

Shiv Rao is a cardiologist and the founder of Abridge, a company building vertical AI systems that turn the conversation between doctor and patient into a stream of notes, orders, prior-authorization justifications, and patient-friendly summaries - giving clinicians back hours each day and letting them look patients in the eye instead of at a screen. Rao argues that in a healthcare system where doctors "need 30 hours a day" and rural hospitals are closing, AI will become the intake coordinator, note-taker, and back-office clerk that makes care accessible and affordable. He is candid that for many routine questions, he would rather his own family start with a capable AI system than a rushed GP - citing studies showing models now give more consistent answers and higher-rated bedside manner than average clinicians - but frames this as an upgrade to the human relationship, where both the patient and the physician arrive at the encounter better prepared. In a newsletter cycle that documents brain implants decoding memory states in real time, mitochondria slipping past cellular defenses in molecular disguise, and industrial infrastructure pivoting to meet the physical demands of the intelligence era, Rao's work illustrates the same principle at the scale of a single clinical encounter: the transition is most powerful when it amplifies what was always there - the human capacity to heal - by removing the barriers that prevented it from doing so.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: the FDA granting Breakthrough Device Designation to an AI-guided brain implant that reads memory failure across 60 neural channels and delivers targeted stimulation at precisely the moment encoding breaks down, a mitochondrial disguise technique achieving 80% cellular delivery efficiency where previous methods barely reached 5%, Google confirming that AI agents now write the substantial majority of its code while engineers migrate from programming to architectural judgment and agent oversight, and a major EV battery factory retooling for grid-scale LFP storage because the physical infrastructure demands of the intelligence era are redirecting industrial capacity faster than consumer markets can absorb it. There's also friction, and it's intense - the Pentagon filing a 40-page legal brief arguing that safety commitments embedded in an AI system's architecture constitute an "unacceptable risk to national security," a rogue AI agent inside Meta exposing sensitive company and user data for two hours because the institution had deployed autonomous capability without building the permission infrastructure to govern it, and BMG's copyright lawsuit against Anthropic extending the collision between AI training practices and the creative economy into the specific songs that define cultural memory. But friction generates texture, and texture is what makes it possible to get a grip on something that would otherwise slip past. Step back for a moment and you can see it: therapeutic systems reaching the resolution where the brain's own mechanisms can be completed rather than bypassed, industrial chemistry pivoting at continental scale toward the substrate the intelligence era requires, the legal system being asked for the first time to rule on whether the values an AI carries into deployment are a design feature or a national security liability, and the software profession reorganizing around a question it has never faced before - what it means to build when the act of building has been delegated. Every transformation has a breaking point. A crucible can destroy what it contains by the intensity of what it applies... or purify it into the only form capable of doing what the moment demands.

AI Releases & Advancements

New today

  • MiniMax: Released MiniMax M2.7, a self-evolving agentic model scoring 56.22% on SWE-Bench Pro and 57.0% on Terminal Bench 2, with parity to Claude Sonnet 4.6 on OpenClaw benchmarks; features recursive self-improvement across skills, memory, and architecture; available immediately via Ollama cloud, OpenRouter, Vercel, and other platforms at $0.30/$1.20 per 1M input/output tokens. (MiniMax)
  • Xiaomi: Released three MiMo-V2 models on March 18–19: MiMo-V2-Pro, a closed-weight flagship with over 1 trillion parameters and 1M token context at $1/$3 per 1M tokens available via API; MiMo-V2-Omni, a multimodal closed-weight model; and MiMo-V2-TTS, a text-to-speech model. (Gizmochina)
  • Nous Research: Released Hermes Agent v0.3.0 on March 17, incorporating 248 pull requests from 15 contributors in 5 days; new capabilities include real-time streaming across CLI and all platforms, a first-class plugin architecture, live Chrome browser control via CDP, local Whisper-based voice mode, PII redaction, and integrations with Browser Use and IDE tools. (GitHub)
  • LangChain: Open-sourced Open SWE on March 17, a background coding agent framework mirroring internal systems used at Stripe, Ramp, and Coinbase; integrates with Slack, Linear, and GitHub, uses subagents plus middleware, and separates harness, sandbox, invocation, and validation layers. (GitHub)
  • LangChain: Launched LangSmith Sandboxes in Private Preview on March 17, providing microVM-isolated environments for AI agents to safely execute untrusted code, spinnable in a single line of code via the LangSmith SDK. (LangChain Blog)
  • Snowflake: Launched Project SnowWork in research preview on March 18, an autonomous enterprise AI platform that orchestrates planning, analysis, and multi-step task execution directly on Snowflake data; available to a limited set of customers. (Snowflake)
  • Amazon: Launched Alexa+ in the UK on March 19 via an early access program tied to new Echo device purchases; the first international rollout of the AI-powered conversational assistant outside North America, with Prime subscribers receiving access for free and non-Prime at £19.99/month once the early access period ends. (TechCrunch)
  • Multiverse Computing: Launched the CompactifAI app and developer API portal, enabling offline inference of compressed models from OpenAI, Meta, DeepSeek, and Mistral AI on edge devices with no cloud dependency; the API portal also added Nemotron 3 family models. (TechCrunch)

Other recent releases

  • OpenAI: Released GPT-5.4 mini and GPT-5.4 nano - smaller, faster, cheaper models optimized for high-volume and latency-sensitive tasks; mini replaces GPT-4.1 mini as the default in ChatGPT and the API, nano is OpenAI's fastest and cheapest model ever. (OpenAI)
  • OpenAI: Launched subagents in Codex - agents can now spin up child agents to handle subtasks in parallel, enabling multi-step autonomous workflows within the coding environment. (Greg Brockman)
  • Tri Dao / Albert Gu: Released Mamba 3, an open-source state-space model that matches or exceeds Transformer performance at equivalent scale while scaling linearly with sequence length instead of quadratically. (Tri Dao)
  • Unsloth: Launched Unsloth Studio, a visual interface for fine-tuning and merging LLMs with no-code workflows, dataset curation, and one-click export to GGUF/ONNX formats. (Unsloth)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more