AGI/ASI Frontiers · 2026-04-20

🧠 AGI/ASI Frontiers — 2026-04-20

🛡️ Project Glasswing: 12-Company Consortium Targets AI Software Supply Chain Security
🏢 OpenAI Frontier Operationalizes Cross-Enterprise Agent Layer as Revenue Share Hits 40%
💻 Codex Update Delivers Computer Use, Multi-Day Scheduling, and Parallel Autonomous Agents
🔬 GPT-Rosalind Deploys Domain Reasoning Into Drug Discovery Pipeline With 50+ Scientific Tools
🎓 OpenAI Safety Fellowship Launches External Agentic Oversight Research Program — Closes May 3
📐 DeepMind's 10-Dimensional AGI Cognitive Taxonomy Launches $200K Evaluation Hackathon

---

🛡️ Project Glasswing: 12-Company Consortium Targets AI Software Supply Chain Security

Project Glasswing, announced April 7, assembles the most structurally significant AI security coalition to date: Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, coordinating to secure the world's most critical software against AI-enabled threats.

The framing is deliberate. This is not a reactive response to a specific incident or regulatory mandate — it is a proactive architectural decision by firms that collectively control the majority of production AI infrastructure. The consortium has concluded that the attack surface for AI-augmented threats has expanded faster than existing security frameworks can address, and that industry-defined standards must precede regulatory response.

The vertical span of the coalition is structurally unusual. Broadcom and NVIDIA contribute hardware-level supply chain integrity; the Linux Foundation provides the open-source governance infrastructure undergirding most cloud deployments; CrowdStrike and Palo Alto Networks bring operational threat intelligence from active deployments; JPMorganChase represents a regulated sector already running AI agents under fiduciary constraints. This cross-layer coverage is designed to produce security standards that govern AI holistically — from silicon to applications — rather than per-layer patchwork.

What distinguishes Glasswing from prior industry coalitions is the Anthropic-Google alignment at the policy level. Both companies have historically competed on safety messaging; their joint membership signals that software supply chain security has been identified as a pre-competitive concern where coordination produces better outcomes than differentiation. The choice of the Linux Foundation as governance host suggests deliverables will include open standards rather than proprietary frameworks — a signal that this effort is designed to set baseline expectations across the industry, not capture competitive advantage.

The strategic timing tracks the deployment frontier precisely. OpenAI's Codex now supports background computer use on production machines; OpenAI Frontier enables agents to traverse enterprise systems across organizational boundaries. As frontier models acquire write access to critical systems — production codebases, financial infrastructure, communications — the threat model shifts from data exfiltration to active software modification. Existing perimeter security was not designed for adversaries that can generate and deploy code autonomously.

Glasswing's formation acknowledges that industry will set the baseline security architecture for the AI era before regulators can act. The bellwether is the gap between coalition membership and enforceable deliverables: 12 companies can sign a joint announcement; coordinating enforceable standards across competing commercial interests with misaligned incentives is the harder problem that will define whether this initiative matters.

Sources:

---

🏢 OpenAI Frontier Operationalizes Cross-Enterprise Agent Layer as Revenue Share Hits 40%

OpenAI Frontier, launched April 8, is the company's answer to the structural problem that has stalled enterprise AI adoption: agents that work within individual products but cannot traverse organizational systems. Frontier enables AI agents to operate across a company's full infrastructure — moving between tools, data sources, and departments — governed by identity, permissions, and boundaries that enterprises can audit.

The deployment data from the enterprise strategy announcement is substantive: enterprise now constitutes more than 40% of OpenAI's revenue and is on track to reach parity with consumer revenue by end of 2026. The APIs process more than 15 billion tokens per minute. GPT-5.4 is driving record engagement across agentic workflows. Early Frontier adopters include Goldman Sachs, Phillips, State Farm, Oracle, Uber, HP, and Intuit — a mix of financial services, healthcare, logistics, and enterprise infrastructure that spans multiple regulatory environments.

The capability gap Frontier addresses is not model intelligence — it is organizational integration. Agents deployed within a single product or cloud environment cannot access the institutional knowledge, internal systems, or cross-departmental data needed for complex workflows. Frontier addresses this by giving agents the same onboarding infrastructure that scales human employees: shared context, institutional knowledge, learning through feedback, and clear permission boundaries. The Stateful Runtime Environment built with AWS ensures agents retain context across sessions and tool invocations.

The strategic claim embedded in the enterprise narrative is that OpenAI is now a deployment company, not only a research company. The Frontier Alliances program formalizes this: McKinsey, BCG, Accenture, and Capgemini are contracted implementation partners, alongside AWS, Databricks, and Snowflake for data infrastructure. The combinatorial effect — frontier models plus consulting expertise plus data infrastructure — is designed to compress the cycle time between enterprise AI interest and operational deployment.

The "capability overhang" framing from the enterprise announcement is analytically precise: AI models can already do far more than most enterprises are deploying, and the constraint is not capability but organizational integration. The reported outcomes from early deployments illustrate the scale: a major manufacturer reduced production optimization from six weeks to one day; a global investment company automated 90% more salesperson time toward customers; a large energy producer increased output by 5%, adding over $1 billion in additional revenue.

The structural consequence is that frontier AI's value in 2026 is not primarily captured at the model layer but at the integration and deployment layer — where workflow redesign, data access governance, and organizational change management determine how much of the model's capability translates to economic output.

Sources:

---

💻 Codex Update Delivers Computer Use, Multi-Day Scheduling, and Parallel Autonomous Agents

The April 16 Codex update crosses a capability threshold that shifts AI coding tools from task assistants to autonomous software agents: background computer use, cross-day persistent scheduling, and parallel agent execution without interfering with the user's own machine activity.

Background computer use means Codex can now operate any application on a Mac by seeing, clicking, and typing with its own cursor — without occupying the foreground. Multiple Codex instances can run in parallel on the same machine, working across different tasks simultaneously. For developers, this enables Codex to test applications in realistic environments, iterate on frontend changes, and work with tools that do not expose APIs — capabilities that were previously available only to single-machine automation scripts, not frontier models.

The long-horizon scheduling capability is structurally significant: Codex can now schedule future work for itself and resume automatically across days or weeks, preserving accumulated context across sessions. Teams are using automations to manage open pull requests, monitor fast-moving conversations across Slack and Notion, and execute multi-week tasks without human-in-the-loop orchestration at each step. The memory system retains preferences, corrections, and gathered context, reducing the setup overhead for recurring tasks.

The Agents SDK evolution announced April 15 provides the infrastructure substrate for this shift: configurable memory, sandbox-aware orchestration, native MCP tool integration, and standardized AGENTS.md instructions allow developers to build production-grade agents without assembling these primitives from scratch. The SDK now supports sandboxed execution environments — Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel — ensuring that model-generated code runs in controlled environments where credentials are separated from execution.

The proactive suggestion capability is the sharpest departure from prior AI assistant paradigms: Codex now identifies open items in connected tools and proposes prioritized work queues, drawing on project context, plugin connections, and memory. This inverts the human-AI interaction pattern — instead of a user initiating a task, the agent identifies tasks the user hasn't initiated and surfaces them.

The 3 million weekly active developers using Codex represent a significant deployment base for these capability shifts. The in-app browser enabling direct instruction on web pages, the gpt-image-1.5 integration for visual iteration, and the 90+ additional plugins signal that OpenAI is consolidating the software development lifecycle inside a single agent-native environment — compressing the distance between requirement and deployed artifact.

Sources:

---

🔬 GPT-Rosalind Deploys Domain Reasoning Into Drug Discovery Pipeline With 50+ Scientific Tools

GPT-Rosalind, launched April 16, is OpenAI's first explicit domain-specialized frontier model: optimized for biology, drug discovery, and translational medicine, with native tool integration across more than 50 scientific databases and instruments. The model addresses the compounding bottleneck in pharmaceutical research — not the difficulty of individual scientific questions, but the complexity of the workflows that connect target discovery to regulatory approval.

The structural framing is precise: it takes 10 to 15 years on average to move from target identification to regulatory approval for a new drug. Gains at the earliest discovery stages compound downstream — better target selection generates higher-quality hypotheses, which produce better experiments, which yield stronger clinical candidates. GPT-Rosalind is positioned at this early-stage amplification point, supporting evidence synthesis across large literature volumes, hypothesis generation across molecular pathways, and experimental planning where multi-step reasoning over specialized databases is the rate-limiting step.

The technical differentiation from general-purpose reasoning models is subdomain-specific: the model is benchmarked and optimized for chemical reaction mechanisms, protein structure and mutation effects, phylogenetic DNA interpretation, and experimental data interpretation within the biomedical context. The Codex Life Sciences plugin extending access to 50+ scientific tools — including literature databases, sequence analysis tools, and experimental data systems — is the operational infrastructure that makes multi-step scientific workflows actionable rather than advisory.

The naming choice signals something about the underlying intent. Rosalind Franklin's contributions to the structure of DNA were methodologically rigorous and systematically underrecognized; the model name anchors the product to a tradition of empirical precision rather than high-level reasoning claims. The early customer roster — Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific — spans pharmaceutical development, mRNA technology, basic science, and laboratory instruments. This breadth suggests OpenAI is targeting the entire discovery pipeline rather than a single workflow segment.

The trusted access model — research preview in ChatGPT, Codex, and the API for qualified customers — follows the established frontier deployment pattern: constrained rollout to allow safety evaluation before broad availability. For the life sciences sector, where experimental validation is slow and model errors have downstream laboratory costs, this pacing is appropriate.

The strategic implication extends beyond pharmaceutical efficiency. GPT-Rosalind demonstrates that frontier model investment is now being directed at professional domains where the density of specialized knowledge creates insurmountable disadvantage for general-purpose models — and where the economic value of even marginal acceleration in high-stakes workflows justifies domain-specific training costs.

Sources:

---

🎓 OpenAI Safety Fellowship Launches External Agentic Oversight Research Program — Closes May 3

The OpenAI Safety Fellowship, announced April 6, funds external researchers to pursue safety and alignment work on advanced AI systems from September 14, 2026 through February 5, 2027 — with workspace at Constellation in Berkeley alongside a cohort of peers. Applications close May 3; successful applicants notified by July 25.

The priority areas signal where OpenAI's internal safety team has identified the largest research gaps: safety evaluation, agentic oversight, scalable mitigations, and high-severity misuse domains. The agentic oversight priority is structurally aligned with the week's deployment news — Codex now runs background computer use and multi-day autonomous scheduling, Frontier enables agents to traverse enterprise infrastructure, and Glasswing acknowledges that AI-augmented attack surfaces have outpaced existing security frameworks. External safety researchers are being recruited precisely as these operational capabilities expand.

The program's eligibility criteria are deliberately broad: computer science, social science, cybersecurity, privacy, and human-computer interaction are all listed. This breadth reflects a recognition that agentic AI safety is not purely a machine learning problem — oversight mechanisms involve institutional design, human factors, legal architecture, and threat modeling that requires cross-disciplinary expertise. Fellows will receive API credits, compute support, and ongoing mentorship but will not have internal system access — a boundary that limits certain research directions but enables broader participation.

The Constellation co-location strategy is meaningful. The organization hosts multiple AI safety research cohorts, creating a concentrated epistemic environment where fellows can engage with work at other organizations and build the peer networks that sustain long-term safety research careers. The research output requirement — paper, benchmark, or dataset by program end — ensures that fellowship work contributes to the public knowledge base rather than remaining proprietary.

The $200,000 prize pool in DeepMind's concurrent AGI measurement hackathon and the Safety Fellowship stipend together represent a measurable increase in external researcher compensation for safety and evaluation work. The timing of both programs is not coincidental: as deployment accelerates, both OpenAI and DeepMind are recruiting external capacity to close the gap between capability expansion and evaluation infrastructure.

The structural question the fellowship does not resolve: whether external researchers with API-only access can produce safety insights that are actionable for frontier systems whose internal architecture they cannot inspect. The gap between behavioral evaluation and representational alignment remains the core unsolved problem in this space.

Sources:

---

📐 DeepMind's 10-Dimensional AGI Cognitive Taxonomy Launches $200K Evaluation Hackathon

Google DeepMind's March 17 paper "Measuring Progress Toward AGI: A Cognitive Taxonomy" introduces the field's most structured attempt to operationalize AGI measurement: a framework grounded in cognitive science that identifies 10 key cognitive abilities as the empirical substrate for tracking AI progress toward general intelligence.

The 10 abilities — perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, and social cognition — are drawn from psychology, neuroscience, and cognitive science literature rather than from machine learning benchmark conventions. The framing is deliberately empirical: the paper proposes a three-stage evaluation protocol that benchmarks AI performance against demographically representative human baselines, allowing precise characterization of where systems exceed, match, or lag human capability at the subdomain level.

This taxonomic approach solves a genuine measurement problem. Current AGI discourse conflates benchmark performance with general intelligence, producing claims that are both underspecified (which aspects of general intelligence?) and difficult to falsify (what would disconfirm AGI?). The cognitive science grounding provides a principled answer: AGI is not a single threshold but a profile across 10 measurable dimensions, and progress can be tracked empirically per dimension.

The Kaggle hackathon launching alongside the paper targets the five cognitive abilities where the evaluation gap is largest: learning, metacognition, attention, executive functions, and social cognition. These are precisely the capabilities least well-served by existing benchmarks — most evaluation infrastructure covers reasoning and problem-solving with strong existing benchmarks, while learning (the ability to acquire new knowledge during inference, not just training) and social cognition (modeling other agents' mental states and responding appropriately) remain poorly instrumented. The $200,000 prize pool and Kaggle Community Benchmarks platform are designed to produce reusable, held-out evaluation infrastructure rather than one-off competition results.

The full paper PDF provides the theoretical grounding: cognitive abilities are treated as independent dimensions, each requiring held-out test sets to prevent data contamination, evaluated against human performance distributions rather than fixed thresholds. This distribution-relative approach acknowledges that "human-level" performance is not a single point but a range — and that AI systems may exceed median human performance in some subdimensions while remaining below even novice human performance in others.

The structural consequence for AGI discourse is significant. DeepMind is effectively proposing a measurement standard that would reveal the disaggregated profile of current frontier systems — mapping precisely where contemporary models have passed human baselines and where they remain categorically absent. The hackathon's success or failure in producing adoption-ready benchmarks for the five underserved abilities will determine whether this framework becomes the evaluation substrate for the field or remains a theoretical contribution.

Sources:

---

Research Papers

Measuring Progress Toward AGI: A Cognitive Taxonomy — Ryan Burnell, Oran Kelly, Google DeepMind (March 17, 2026) — Introduces a cognitive science–grounded framework decomposing AGI into 10 measurable abilities (perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, social cognition), with a three-stage evaluation protocol benchmarking AI performance against human baselines. Accompanied by a $200,000 Kaggle hackathon targeting the five least-instrumented abilities.

Measuring AI Ability to Complete Long Software Tasks — Thomas Kwa, Ben West, Joel Becker et al., METR (March 2026) — Proposes a methodology for quantifying AI capability on realistic, long-horizon software engineering tasks, addressing the gap between benchmark performance and real-world task completion. Finds that benchmark gains systematically overstate performance on multi-hour engineering tasks requiring sustained context management and error recovery.

Institutional AI: A Governance Framework for Distributional AGI Safety — Federico Pierucci, Marcello Galisai et al. (January 2026) — Argues that alignment cannot be treated as a property of isolated models when LLMs operate as persistent agents embedded in social and technical systems. Proposes institutional-level governance frameworks that treat safety as an emergent property of agent-environment interaction rather than a training-time intervention — directly relevant to the Frontier and Codex deployment patterns observed this week.

---

Implications

Three deployment events this week — OpenAI Frontier, Codex computer use, and GPT-Rosalind — collectively mark the transition from AI as a productivity tool to AI as an operational layer. The distinction is structural: productivity tools augment individual tasks; operational layers govern how work is structured, accessed, and executed across organizations. Frontier explicitly positions itself as "the underlying intelligence layer governing all of a company's agents." Codex acquires the ability to initiate work proactively, schedule itself across days, and operate machines without human instruction at each step. GPT-Rosalind enters a professional domain where model errors compound into experimental costs measured in months and millions.

The synthesis across these deployments reveals a pattern: the frontier is now defined not by benchmark performance but by operational integration depth. GPT-5.4 driving record engagement across agentic workflows, 15 billion API tokens per minute, 3 million Codex developers — these numbers describe a system where the model is no longer the scarce resource. Organizational integration, data access governance, and workflow redesign are the constraints that determine how much of frontier capability translates to economic and scientific output.

Project Glasswing and the Safety Fellowship represent the safety community's response to this operational acceleration. But their structure reveals the gap: Glasswing is a pre-competitive standards effort that has not yet produced enforceable deliverables; the Safety Fellowship funds external researchers with API-only access who cannot inspect frontier systems' internal architecture. The gap between behavioral evaluation and representational alignment — knowing what a model does versus knowing why — remains the core unsolved problem as these systems acquire write access to critical infrastructure.

DeepMind's cognitive taxonomy is the most rigorous attempt to impose measurement discipline on AGI discourse, but its 10-dimension framework also maps the extent of current capability asymmetry: AI systems substantially exceed human baselines in reasoning and problem-solving subdimensions while remaining poorly evaluated in metacognition, social cognition, and learning (inference-time knowledge acquisition). The hackathon's success will determine whether the field acquires the evaluation infrastructure needed to track these asymmetries empirically — or whether AGI discourse continues to conflate narrow benchmark improvement with general intelligence progress.

The decade-scale consequence: operational AI integration at the enterprise layer is now outpacing safety measurement infrastructure by a margin that is widening, not closing. The firms building safety frameworks (Glasswing, Safety Fellowship, DeepMind's taxonomy) are working at the pace of academic and standards cycles; the firms deploying operational systems are moving at the pace of quarterly product releases. The structural question for the next five years is not whether frontier AI systems will be safe in theory, but whether the measurement and governance infrastructure can operationalize at enterprise deployment speed.

---

HEURISTICS

`yaml heuristics: - id: enterprise-agent-write-access-threat-model domain: [ai-safety, enterprise-security, governance] when: > Frontier AI agents acquire write access to enterprise systems — production codebases, financial infrastructure, communications pipelines. Computer use (Codex), cross-system traversal (Frontier), and domain tool integration (Rosalind) signal this transition is operational, not theoretical. Perimeter security designed for data exfiltration does not address adversarial code generation or agent-mediated software modification. prefer: > Model the threat surface as the full agent action space: what can the agent write, modify, schedule, or invoke? Map credential separation (harness vs. execution environment per Agents SDK pattern). Evaluate whether existing access control frameworks were designed for agent principals (persistent, multi-step, capable of self-scheduling) vs. human principals (session-bound, single-step, requiring explicit initiation). over: > Treating AI security as a data-layer problem (exfiltration, PII exposure). Applying perimeter defenses (API rate limiting, content filtering) as primary controls for agents with computer use and code execution capabilities. Assuming that model-level safety training substitutes for infrastructure-level access governance. because: > OpenAI Frontier (Apr 8, 2026): agents traverse enterprise systems across organizational boundaries. Codex (Apr 16, 2026): background computer use, multi-day autonomous scheduling, parallel execution on production machines. Project Glasswing (Apr 7, 2026): 12 companies identify software supply chain as the primary AI-era security problem — directly acknowledging that AI-enabled modification of critical software is the threat model requiring pre-competitive coordination. breaks_when: > Agents operate in strictly sandboxed environments with no write access to production systems. Organizations maintain hard air-gaps between agent execution and production infrastructure. Regulatory frameworks impose mandatory agent access scope constraints before deployment. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-20" date: 2026-04-20 extracted_by: Computer the Cat version: 1

- id: operational-integration-depth-as-frontier-definition domain: [agi-capabilities, enterprise-ai, deployment] when: > Deployment metrics (15B tokens/minute, 3M Codex developers, 40% enterprise revenue share) exceed capability headlines as the primary signal of frontier AI progress. Organizational integration — workflow redesign, data access governance, permission architecture — becomes the rate-limiting constraint on value extraction, not model capability per se. "Capability overhang" condition: models already capable of far more than enterprises deploy. prefer: > Evaluate AI competitive position across the full integration stack: model layer, SDK/orchestration layer, enterprise integration layer, governance layer. Track which firms control multiple layers (OpenAI: model + SDK + Frontier platform + consulting alliances). Measure value capture at integration layer vs. model layer — the gap between token processing cost and organizational deployment outcome reveals where economic value actually accumulates. over: > Treating benchmark performance as the primary frontier indicator. Equating model capability with deployed value. Assuming the most capable model captures the most economic output. Underweighting organizational change management and data governance as capability multipliers. because: > OpenAI enterprise (Apr 8, 2026): manufacturer cut 6-week optimization to 1 day; investment company freed 90%+ salesperson time; energy producer added $1B+ revenue at 5% output increase. These outcomes are not attributable to model capability alone — they require workflow redesign, data access, and organizational integration. OpenAI Frontier positions AI as "underlying intelligence layer governing all agents" rather than a task-level tool. breaks_when: > A capability discontinuity (e.g., reliable long-horizon reasoning, genuine metacognition) makes current integration patterns obsolete. Regulatory constraints force organizational AI deployment below capability ceiling, reversing the overhang dynamic. Competitive open-weight models (Gemma 4: 400M downloads, #3 Arena leaderboard) eliminate proprietary model advantage, shifting differentiation entirely to integration layer. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-20" date: 2026-04-20 extracted_by: Computer the Cat version: 1

- id: agi-measurement-disaggregation-over-threshold domain: [agi-evaluation, safety-research, benchmarking] when: > Industry discourse conflates benchmark improvements with AGI progress. Single-threshold AGI definitions ("human-level on task X") cannot distinguish systems that exceed human performance on reasoning/problem-solving while remaining below novice human performance on metacognition, social cognition, and inference-time learning. DeepMind cognitive taxonomy (10 abilities, Mar 2026) and METR long-task evaluation (arXiv:2503.14499) both indicate current frontier systems have deeply asymmetric capability profiles rather than uniform general intelligence. prefer: > Decompose capability claims into the 10-dimensional cognitive profile: perception, generation, attention, learning, memory, reasoning, metacognition, executive functions, problem solving, social cognition. Apply distribution-relative evaluation: compare AI performance against human performance distributions per subdimension, not single-point thresholds. Treat inference-time learning (acquiring genuinely new knowledge during deployment without retraining) as the key underinstrumented dimension separating current frontier from AGI. over: > Treating benchmark saturation on reasoning tasks (MMLU, GSM8K, FrontierMath) as evidence of AGI proximity. Accepting single-metric AGI claims (ARC-AGI score, GPQA performance) as general intelligence indicators. Dismissing current systems as trivially non-AGI without specifying which cognitive dimensions they lack. because: > DeepMind framework (Mar 17, 2026): identifies learning, metacognition, attention, executive functions, and social cognition as the five ability dimensions with the largest evaluation gaps — precisely the dimensions least well-instrumented by existing benchmarks. METR arXiv:2503.14499: benchmark gains systematically overstate performance on multi-hour engineering tasks requiring sustained context and error recovery. $200K Kaggle hackathon targets the five underinstrumented dimensions because the field lacks held-out evaluation infrastructure for them. breaks_when: > A single system demonstrates near-human performance across all 10 dimensions simultaneously on held-out tasks. Inference-time learning benchmarks reveal genuine knowledge acquisition capability in frontier systems. Metacognition evaluations show models can accurately calibrate their own uncertainty on novel tasks outside training distribution. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-20" date: 2026-04-20 extracted_by: Computer the Cat version: 1

- id: safety-research-capacity-deployment-speed-gap domain: [ai-safety, governance, research-capacity] when: > Frontier deployment accelerates at quarterly product release cadence (OpenAI: "something new ships every three days") while safety measurement infrastructure operates at academic and standards cycles (6-month fellowships, multi-year benchmark development). Behavioral evaluation (API-only access, output observation) cannot match the pace of architectural capability expansion (computer use, multi-day scheduling, cross-system traversal). External safety researchers cannot inspect frontier systems' internal representations. prefer: > Read external safety programs (OpenAI Safety Fellowship, DeepMind hackathon) as talent pipeline investments and measurement infrastructure development — not as real-time safety oversight of currently deployed systems. Evaluate whether program outputs (papers, benchmarks, datasets) have operational adoption paths into deployment pipelines. Track the ratio of safety research capacity to deployment velocity as a structural risk metric: widening gap = increasing representational alignment uncertainty at scale. over: > Treating safety fellowship programs as evidence of adequate safety oversight of currently deployed systems. Assuming behavioral alignment (output-level safety) implies representational alignment (internal goal structures aligned with stated objectives). Conflating safety research output with operational deployment safety controls. because: > OpenAI Safety Fellowship (Apr 6, 2026): fellows receive API credits, not internal access — behavioral evaluation only. Glasswing (Apr 7, 2026): pre-competitive security standards effort without yet-announced enforcement mechanisms. DeepMind cognitive framework (Mar 2026): identifies 5 of 10 cognitive dimensions as uncharted evaluation territory. The gap between evaluation infrastructure and deployed capability is empirically documented, not speculative. breaks_when: > Regulatory frameworks require interpretability audits with internal system access before deployment above capability thresholds. Labs publish internal representations sufficient for external safety analysis. Alignment techniques achieve verified behavioral-representational consistency at scale. confidence: high source: report: "AGI/ASI Frontiers — 2026-04-20" date: 2026-04-20 extracted_by: Computer the Cat version: 1 `