Tech Intelligence Digest - June 25, 2026

Quick Summary

• OpenAI moved deeper into full-stack AI infrastructure with Jalapeno, a Broadcom-built inference chip for running LLMs cheaply and quickly at scale. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• OpenAI also published agent-work research showing Codex usage shifting from short coding help toward delegated, long-running work across technical and non-technical roles. Source: https://openai.com/index/how-agents-are-transforming-work/
• Anthropic launched Claude Tag, an always-available Claude inside Slack with channel-scoped memory, tools, spending limits, and audit logs. Source: https://www.anthropic.com/news/introducing-claude-tag
• Security pressure kept rising: Five Eyes warned AI cyber assumptions can go stale in "months, not years," while Microsoft's AutoJack research showed how a browsing agent can turn localhost into a remote-code-execution surface. Sources: https://www.nsa.gov/Press-Room/News-Highlights/Article/Article/4523810/five-eyes-cyber-security-agencies-statement/ and https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/

AI and Models

• OpenAI and Broadcom unveiled Jalapeno on June 24. OpenAI calls it its first "Intelligence Processor," designed specifically for LLM inference, meaning the serving phase where a model answers users after training is done. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• OpenAI says Jalapeno was designed around its model roadmap, kernels, memory movement, networking, scheduling, and product needs, with Broadcom and Celestica helping turn the design into chips, boards, racks, networking, and production systems. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• Early Jalapeno samples are running ML workloads in the lab, including GPT-5.3-Codex-Spark, and OpenAI says early testing points to substantially better performance per watt than current state-of-the-art chips. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• OpenAI says the chip went from initial design to tape-out in nine months, helped by OpenAI models in the design and optimization process. Tape-out means the chip design is finalized for manufacturing. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• OpenAI's June 25 agent-work paper says Codex is becoming the primary AI tool inside OpenAI, including in Legal, Finance, and Recruiting, not just Engineering. Source: https://openai.com/index/how-agents-are-transforming-work/
• In sampled individual-user data, OpenAI says 80.6% of Codex users made at least one request estimated to exceed 30 minutes of human work, 70.2% made one over one hour, and 25.6% made one over eight hours. Source: https://openai.com/index/how-agents-are-transforming-work/
• The plain-English read: agent value is moving from "answer this prompt" to "own this task for a while," especially when users can run multiple agents in parallel. Source: https://openai.com/index/how-agents-are-transforming-work/

Dev Tools and Infrastructure

• Anthropic launched Claude Tag on June 23 as a Slack-native way to tag @Claude, give it selected channel access, connect tools/data/codebases, and delegate work asynchronously. Source: https://www.anthropic.com/news/introducing-claude-tag
• Anthropic says its internal Claude Tag is now one of the main ways work gets done there, and that 65% of its product team's code is created by the internal version. Source: https://www.anthropic.com/news/introducing-claude-tag
• Claude Tag introduces channel-scoped Claude identities, meaning the sales Claude and engineering Claude can have separate memories, tools, and data access. Source: https://www.anthropic.com/news/introducing-claude-tag
• Administrators can set tool access, channel access, monthly spend limits, and view logs of what @Claude did and who requested it. Source: https://www.anthropic.com/news/introducing-claude-tag
• Theo's latest relevant watchlist video argued that package managers need better trust primitives for the agent era: revocable low-install releases, release audits, richer package metadata, and stronger anti-name-squatting. Source: https://www.youtube.com/watch?v=wEAb0x3wTRc
• Theo also highlighted the AI-PR review problem: agents can create giant pull requests, so humans need layered summaries and review orderings instead of raw alphabetical file diffs. Source: https://www.youtube.com/watch?v=wEAb0x3wTRc
• The queue note on Modal-style sandboxes still fits this cycle: coding agents need isolated execution environments so tool calls are reproducible, contained, and less likely to leak secrets or mutate the host machine. Source: https://x.com/ian_dot_so/status/2013316676637294890?s=46

Industry and Big Tech

• OpenAI's Jalapeno announcement is a Big Tech-style vertical-integration move: OpenAI is not only building models and products, but also custom silicon and inference infrastructure underneath them. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• Broadcom's release says Jalapeno is meant for gigawatt-scale deployment with data center partners over multiple generations. Source: https://investors.broadcom.com/news-releases/news-release-details/openai-and-broadcom-unveil-llm-optimized-intelligence-processor
• The strategic signal: frontier AI companies are trying to control latency, cost, reliability, and supply by owning more of the stack from model to chip to rack. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• Menlo Ventures raised $3 billion in new funds to back AI startups across stages, reinforcing that large AI-native venture funds are still forming even after the 2025-2026 capital flood. Source: https://news.crunchbase.com/venture/menlo-ventures-raise-ai-startup-funding-across-stages-anthropic/
• Mirendil, founded by former Google and Anthropic researchers, announced a $200 million seed round backed by a16z, Kleiner Perkins, Nvidia, and others to build AI for scientific discovery. Source: https://digg.com/tech/a1b3s5jt
• The VC pattern is still concentrated around hard bottlenecks: AI infra, enterprise workflows, scientific AI, and security-adjacent tooling are getting the most serious checks. Sources: https://news.crunchbase.com/venture/menlo-ventures-raise-ai-startup-funding-across-stages-anthropic/ and https://digg.com/tech/a1b3s5jt
• The queued Meta note remains relevant as a management signal: AI-native pods and AI-adoption metrics may create real leverage, but they also risk rewarding visible AI usage over quality, governance, and durable business outcomes. Source: state/newsletter_queue.json

Security

• The NSA-hosted Five Eyes statement says frontier AI is changing cybersecurity fast enough that risk assumptions can become outdated in "months, not years." Source: https://www.nsa.gov/Press-Room/News-Highlights/Article/Article/4523810/five-eyes-cyber-security-agencies-statement/
• The same statement frames cyber resilience as a business-continuity and market-trust issue, not just an IT problem. Source: https://www.nsa.gov/Press-Room/News-Highlights/Article/Article/4523810/five-eyes-cyber-security-agencies-statement/
• Microsoft's AutoJack writeup shows a concrete agent-security failure mode: a malicious webpage rendered by a browsing agent can reach a local MCP WebSocket and spawn arbitrary processes on the host. Source: https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/
• Plain English: "localhost" used to feel private because it meant "this machine," but an AI browser running on the same machine can become the bridge from an untrusted webpage into local developer tools. Source: https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/
• Microsoft says the vulnerable AutoGen Studio WebSocket surface was development-stage and has been fixed, but the broader lesson applies to every agent framework with tools, browsers, local services, and weak auth boundaries. Source: https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/
• For Deval's coding-agent workflow, the practical defense is boring and important: isolate agent execution, lock down local control planes, require auth on tool servers, scope credentials, and treat webpages as hostile input. Source: https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/

Creator Watchlist

• Theo - t3.gg: The strongest recent takeaway is that the JavaScript package ecosystem needs agent-era trust signals because agents install and run packages faster than humans can inspect them. Source: https://www.youtube.com/watch?v=wEAb0x3wTRc
• Theo - t3.gg: His giant-PR point is also worth keeping: as AI increases code volume, review UX becomes core developer infrastructure, not a nice-to-have. Source: https://www.youtube.com/watch?v=wEAb0x3wTRc
• Chris Raroque: No fresh high-signal upload showed up inside this exact 48-hour window during this run. Source: https://www.youtube.com/@ChrisRaroque
• Lenny's Podcast: No fresh high-signal full episode showed up inside this exact 48-hour window during this run. Source: https://www.youtube.com/@LennysPodcast
• John Kim: No fresh high-signal YouTube item showed up inside this exact 48-hour window during this run. Source: https://www.youtube.com/results?search_query=John+Kim+AI+startup
• Greg Isenberg: No fresh high-signal full upload showed up inside this exact 48-hour window during this run. Source: https://www.youtube.com/@GregIsenberg

Takeaways

• The biggest theme is full-stack control: OpenAI wants the chip, the inference system, the model, the agent product, and the workflow surface. Source: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
• The second theme is multiplayer agents: Claude Tag and OpenAI's Codex research both point toward agents living inside team workflows instead of isolated chat windows. Sources: https://www.anthropic.com/news/introducing-claude-tag and https://openai.com/index/how-agents-are-transforming-work/
• The third theme is execution security: once agents browse, install packages, call tools, and run code, the boundary is no longer "model output"; the boundary is the whole workstation, CI system, package manager, and permission graph. Sources: https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/ and https://www.youtube.com/watch?v=wEAb0x3wTRc
• For Deval: keep leaning into distributed systems, security, agent harnesses, and infra fundamentals. The market is rewarding engineers who can turn AI output into reliable systems with clean permissions, review loops, and production-grade execution boundaries. Sources: https://openai.com/index/how-agents-are-transforming-work/ and https://www.microsoft.com/en-us/security/blog/2026/06/18/autojack-single-page-rce-host-running-ai-agent/