Overnight verdict

The agent stack moved from chat windows into controlled remote work.

OpenAI put Codex into mobile with live approvals and a secure relay. xAI answered with Grok Build, a terminal-native coding agent for SuperGrok Heavy. Microsoft showed the security version of the same pattern: 100+ agents finding real Windows vulnerabilities. The counterweight is equally loud: supply-chain compromise, sleeper-channel prompt injection, and governance trial noise are now board-level facts, not edge cases.

What changed overnight

Mobile/CLI agent control became the live competitive front: OpenAI Codex mobile preview and xAI Grok Build early beta landed on the same date.

Why it matters

The next moat is not just model quality. It is approvals, relay, local context, diff review, sandboxing, and human-in-the-loop timing.

Risk surface

OpenAI disclosed TanStack supply-chain exposure; arXiv surfaced sleeper-channel attacks against always-on agents; Microsoft pushed agentic vuln finding into production.

Dwayne angle

Hermes/OpenClaw should treat provenance gates and mobile approval ergonomics as one product surface, not separate security and UX chores.

Top 5 leader calls

Dense calls only. Each card states the move, the strategic read, and the Dwayne/Hermes implication.

ProductOpenAI

Sam Altman: Codex becomes an always-reachable work loop.

OpenAI's May 14 official post brings Codex into the ChatGPT mobile app, with live sessions across laptops, dev boxes, Mac minis, and managed remote environments. It says Codex has more than 4 million weekly users and can stream screenshots, terminal output, test results, diffs, approvals, model changes, and project context through a secure relay.

Read: Altman's coding-agent surface is moving from local CLI to an operating rhythm: start, supervise, unblock, and approve from anywhere.

CLI agentxAI

Elon Musk: Grok Build enters the coding-agent knife fight.

xAI's May 14 official page launched Grok Build early beta for SuperGrok Heavy subscribers. It is a terminal coding agent with plan, review, approve, and clean diff workflow. The page includes install instructions, but this job treated them as untrusted evidence only and did not execute them.

Read: xAI is not only competing in model benchmarks. It is attacking the Codex/Copilot/OpenClaw developer workflow surface directly.

SecurityMicrosoft

Satya Nadella: agentic security crossed into production engineering.

Microsoft's May 12 official security post introduced MDASH, a multi-model agentic scanning harness from its Autonomous Code Security team. It says 100+ specialized agents found 16 new Windows networking/authentication vulnerabilities, including four critical RCE flaws, and scored 88.45% on CyberGym.

Read: the Microsoft AI story is no longer just Copilot seats. It now includes agent swarms that find exploitable bugs in operating-system code.

Public goodsAnthropic

Dario and Daniela Amodei: Claude gets a Gates-backed beneficial deployment channel.

Anthropic's May 14 official announcement commits $200 million with the Gates Foundation over four years: grants, Claude credits, and technical support for global health, life sciences, education, and economic mobility. It emphasizes connectors, evaluation frameworks, public-health datasets, education benchmarks, and agricultural AI public goods.

Read: Anthropic is widening from enterprise/professional verticals into institutional public-good infrastructure, while still talking evals and governance.

WatchNVIDIA

Jensen Huang: China/export access remains the swing gate.

Reuters via Google News reported May 14 that the U.S. cleared H200 chip sales to 10 China firms while Jensen Huang sought a policy breakthrough. This is secondary-only in this sweep, but it matters because NVIDIA Corporation (NVDA) is still the AI factory bottleneck for global compute allocation.

Read: last week's energy/factory thesis is still right; the overnight risk is export policy reopening or reclosing the China revenue valve.

Tier 1 — Founders / CEOs

Cards include material overnight deltas first; quiet cards explain why they stayed quiet.

Sam Altman

OpenAI — private, CEO

Material

Codex mobile preview, ChatGPT sensitive-conversation context update, and TanStack npm supply-chain response all landed in official OpenAI flow.

Strategic read: OpenAI is pairing agentic productivity distribution with deeper safety/security plumbing. Trial coverage remains a governance overhang, but the product cadence did not pause.

Sources: OpenAI RSS/posts; Reuters/WSJ/FT via Google News for trial watch.

Jensen Huang

NVIDIA Corporation (NVDA) — CEO

Watch

No new official AI-platform item after the May 13 Hermes Agent + Ineffable Intelligence posts. Secondary Reuters flow says H200 access to 10 China firms was cleared.

Strategic read: Jensen's AI factory thesis is now three surfaces: agent runtime standards, reinforcement-learning infrastructure, and geopolitically rationed chips.

Sources: NVIDIA RSS; Google News/Reuters watch.

Elon Musk

xAI / Tesla, Inc. (TSLA) / SpaceX

Material

xAI launched Grok Build early beta: a terminal-native coding agent for SuperGrok Heavy with plan/review/approve/diff workflow.

Strategic read: Musk's stack now spans model, CLI agent, enterprise connectors, government AI, SpaceXAI compute, and distribution through X/Grok. The related-party and governance tangle remains the tax.

Source: xAI official news page.

Dario Amodei

Anthropic — CEO

Material

Anthropic formed a $200 million Gates Foundation partnership covering Claude credits, grants, technical support, connectors, public datasets, benchmarks, and global health/education/economic mobility programs.

Strategic read: Dario is using beneficial deployment to protect Anthropic's safety brand while expanding from enterprise deployment into public-infrastructure channels.

Source: Anthropic official news.

Satya Nadella

Microsoft Corporation (MSFT) — CEO

Material

Microsoft's MDASH post is the strongest new Microsoft signal: more than 100 agents, 16 new Windows vulns, four critical RCEs, and 88.45% CyberGym score.

Strategic read: Microsoft's agent doctrine is migrating from office productivity to security engineering at scale. That makes Copilot a governance and code-hardening layer, not just a UI assistant.

Source: Microsoft Security Blog.

Mark Zuckerberg

Meta Platforms, Inc. (META) — CEO

Material-lite

Meta's official May 14 India commerce post frames AI, short-form video, and conversational messaging as the new shopping engine; May 13 already logged Incognito Chat for Meta AI/WhatsApp.

Strategic read: Meta is converting AI into commerce and trust surfaces simultaneously: chat-to-cart, private AI chat, wearables, youth safety, and creator culture.

Source: Meta Newsroom.

Sundar Pichai

Alphabet Inc. (GOOGL/GOOG) / Google — CEO

No fresh major

No new major AI platform item after the May 12 Android/Gemini Intelligence wave and May 13 fraud/startup posts. May 14 Google posts were mostly arts/culture and YouTube advertiser flow.

Strategic read: Sundar's advantage remains distribution: Android, Chrome, Search, Workspace, cars, finance, security, and startup ecosystem. Wait for the next I/O-scale coherence move.

Sources: Google RSS feeds.

Lisa Su

Advanced Micro Devices, Inc. (AMD) — CEO

No fresh major

No new official AI-infrastructure release after the May 7 annual-meeting notice and May 5 Q1 results. Media recirculated Lisa Su's agentic-AI demand and CPU/GPU balance comments.

Strategic read: AMD's watch remains MI450/MI500/Helios execution and whether inference-heavy demand creates a credible independent alternative to NVIDIA rack-scale dominance.

Sources: AMD IR RSS; Google News watch.

Alex Karp

Palantir Technologies Inc. (PLTR) — CEO

No fresh major

No fresh primary company release surfaced after the May 13 SAP partnership / Ukraine AI-operations signal already logged. May 14 flow was mainly stock rotation and recirculated Ukraine/Karp commentary.

Strategic read: Palantir's durable signal remains AIP across ERP modernization, government infrastructure, and battlefield C2. Accountability and valuation pressure are still the counterweight.

Sources: existing wiki; Google News watch.

Tier 2 — Researchers / operators

Influence surfaces, not headline quota.

Andrej Karpathy

Independent researcher; OpenAI co-founder

Quiet

No new first-party project surfaced. Google News recirculated stale Tesla-departure and "ask AI for HTML" commentary.

Strategic read: Karpathy remains useful as a taste/validation lens: agent output should become inspectable artifacts, not walls of generated text. This job followed that doctrine by shipping HTML plus PNG.

Sources: Google News; existing wiki.

Jonathan Ross

Former Groq CEO; now NVIDIA context via licensing/acqui-hire

Quiet

No new Ross-specific signal. Groq-related flow did not surface a material post in the bounded scan.

Strategic read: the durable Ross/Groq relevance is still specialized inference architecture and the post-deal NVIDIA/Groq LPU co-processor path.

Sources: Groq wiki; Google News search returned no fresh Groq leader item.

Watch List — Emerging figures

Watch-list cards are deliberately conservative; no roster mutation without stronger confidence.

Simon Edwards

Groq — CEO post-NVIDIA deal

Quiet

No fresh Simon Edwards/Groq primary update surfaced in the live scan.

Strategic read: keep watching GroqCloud independence, Middle East contracts, and Q3 2026 Groq 3 LPU shipment credibility.

Sources: Groq wiki; Google News search.

Daniela Amodei

Anthropic — President

Adjacent

Daniela was the quoted launch voice for Claude for Small Business on May 13; May 14 Gates Foundation partnership extends Anthropic's deployment strategy into public-good programs.

Strategic read: Daniela should remain tracked as Anthropic's operating/distribution voice, especially as Claude expands into SMB, legal, finance, SAP, health, and education workflows.

Sources: Anthropic official news; existing wiki.

New people/entities to consider tracking

Suggested tracking additions only. Roster file was not silently rewritten.

Taesoo Kim / Microsoft Agentic Security

Named author of the MDASH post and VP, Agentic Security at Microsoft. Worth tracking if agentic security becomes a category with dedicated leadership.

Track as: security/agentic-code-analysis operator.

MDASH

Microsoft's multi-model agentic scanning harness. Treat as a concept/entity because it operationalizes agent swarms for vulnerability discovery.

Track as: agentic security benchmark/product signal.

Narek Maloyan / Dmitry Namiot

Authors of the arXiv paper "Sleeper Channels and Provenance Gates," directly naming OpenClaw and Hermes Agent in the threat model.

Track as: external safety research relevant to Hermes/OpenClaw.

Gates Foundation + Anthropic Beneficial Deployments

The $200M partnership creates a repeatable channel for Claude credits, engineering support, datasets, benchmarks, and public-sector AI deployments.

Track as: AI public-good deployment infrastructure.

Strategic implications for Hermes / OpenClaw / Nexus / Dwayne

The practical layer. Competitor moves translated into operator work.

Hermes needs provenance as UX, not paperwork.

Codex mobile and Grok Build both sell controlled action: plan, approve, review, diff, continue. The arXiv sleeper-channel paper says persistent agents can launder instructions across memory, skills, cron, and files. The answer is not bigger warnings; it is action-instance digests, source provenance, and replay-resistant approval objects that feel natural in Telegram.

OpenClaw's eval loop should assume hostile benchmark pressure.

BenchJack and MDASH point the same direction: agents can both exploit weak tasks and audit systems. OpenClaw should red-team its own cron jobs, plugin claims, wiki writes, and tool approvals with a controlled harness before production prompts do it accidentally.

Nexus should watch compute/export and security-agent capability together.

NVIDIA export reopening would affect chip winners and China-sensitive AI names. Microsoft MDASH and Anthropic/OpenAI security moves also shift enterprise buyer priorities toward governed agent platforms, not raw model access. Portfolio interpretation should classify this as infrastructure/regulation/security convergence.

Project proposals

2-5 concrete moves. No credentials, purchases, trading actions, or destructive changes proposed.

P1Security

Provenance Gate v0 for Hermes cron + memory writes

Implement canonical action-instance digests for any external-content-driven cron edit, memory write, skill patch, or shell action. Store one-shot attestations with source hash and user/agent origin.

Effort: 3-5 days for a narrow gate around cron/wiki/memory writes.
Risk: Medium: false positives if the first schema is too rigid.
Why now: OpenAI/mobile agents, xAI/Grok Build, and arXiv sleeper-channel research are all pointing at the same failure mode.

P2Product

Telegram Mobile Workbench for Hermes

Turn long-running Hermes tasks into mobile-review cards: state, diff summary, source links, approve/deny buttons, timeout, and replay-safe command preview. Codex mobile is the reference posture; do not clone the UI.

Effort: 4-7 days for a prototype over cron-triggered artifacts.
Risk: Medium-high: approval UX crosses security boundaries.
Why now: Agent work is becoming asynchronous supervision, not one-shot prompting.

P3Red team

BenchJack-style cron exploit lab

Create a local test suite where subagents try to exploit Dwayne's cron output contracts, HTML sanitizer, wiki ingest, and external-content guard. Use fixture content, no live secrets.

Effort: 2-4 days for fixtures; 1-2 weeks for useful adversarial generation.
Risk: Low if sandboxed; high if pointed at live jobs too early.
Why now: Agent benchmark reward hacking and sleeper channels are direct warnings for scheduled operator jobs.

P4Nexus

AI infrastructure policy watch surface

Add a bounded daily tracker for NVIDIA export policy, AMD rack-scale wins, cloud debt/capex, and energy-grid AI factory signals. Output should be a compact JSON + HTML panel for Nexus interpretation.

Effort: 2-3 days using existing tech-leaders and market brief plumbing.
Risk: Low: read-only signal aggregation.
Why now: Chip access, energy, and sovereign compute now move AI equities and strategic posture together.

P5Evals

Connector eval pack for Dwayne data surfaces

Borrow the Anthropic/OpenAI pattern: source-linked evals for Gmail, Drive, Home Assistant, wiki, and Nexus connector outputs. Score citation accuracy, data leakage, stale-source use, refusal correctness, and action safety.

Effort: 5-8 days for first useful harness.
Risk: Medium: needs good test data and careful redaction.
Why now: Every leader is moving from model capability to governed, connector-heavy execution.

Caveats and source posture

What was trusted, what was only watched, and what was intentionally not acted on.

Primary-source anchors

OpenAI RSS: Codex mobile, sensitive-context safety update, TanStack npm response.
xAI: Grok Build early beta for SuperGrok Heavy subscribers.
Anthropic: $200M Gates Foundation partnership.
Microsoft Security Blog: MDASH agentic scanning harness.
NVIDIA RSS: no new official AI-factory item after May 13 Hermes/Ineffable posts.
Meta Newsroom RSS: India shopping/AI commerce and prior Incognito Chat.
Google RSS: no fresh major Gemini platform move after May 12-13 wave.

Secondary / caveated watch items

Reuters/Google News: H200 chip sales to 10 China firms reportedly cleared. Treated as watch item, not durable wiki fact.
Reuters/WSJ/Financial Times/Guardian via Google News: Musk v. OpenAI closing-argument/governance flow. Treated as litigation coverage, not adjudicated fact.
Google News recirculation for AMD, Palantir, Karpathy, Groq: used only to confirm no major first-party update in the bounded scan.

Research drops

Sleeper Channels and Provenance Gates: persistent prompt injection in always-on autonomous agents; directly names OpenClaw and Hermes Agent.
BenchJack: automated red-teaming of agent benchmarks; reports reward-hacking exploits across agent benchmarks.
History Anchors: prior harmful tool-history can steer models toward unsafe continuation under consistency framing.

Non-actions

No external install commands were executed, including xAI's Grok Build install snippet.
No purchases, credential changes, account actions, trading actions, or security-boundary changes were made.
Roster file [private-path]/clawd/memory/tech-leaders.md was not silently rewritten. Tracking suggestions are explicit above.