AI Jobs Brief

2026-06-25 · daily AI labor-market brief
📰 News brief · Maya (VibeVoice)

download MP3 · single-anchor news read

🎙 Podcast · Maya + Carter (VibeVoice)

download MP3 · two-host conversation

🥇 Today's top three

  1. The Unfireable Safety Kernel: new research proposes execution-time controls outside the agent's runtime · arXiv
  2. Wired: Anthropic cofounder Dario Amodei replaced by Tom Brown in White House meetings · Wired
  3. A fake AI agent skill bypassed every security scanner and reached ~26,000 agents, including corporate accounts · The Hacker News

1 · AI replacing jobs

Tech layoffs remain severe: a running list at TechCrunch tallies nearly 155,000 jobs cut across major firms including Meta, LinkedIn, Lucid, and Robinhood in 2026, with AI frequently named as the driver. Oracle's 21,000 cuts are now being paired with billions in debt-funded AI data-center buildout — the same dollars are moving from headcount to compute. For early-career workers, the durable answer is the skill stack that AI cannot easily commoditize: AI operations, security, systems thinking, and the judgment to use automation without being replaced by it.

2 · AI hiring & new opportunities

The opportunity is in the controls and infrastructure that make agentic AI safe to deploy. New research on an "unfireable safety kernel" argues runtime controls must live outside the agent's own address space so the agent cannot bypass its own guardrails — a direct hiring signal for AI safety engineering, formal verification, and runtime monitoring roles. On the policy side, the Anthropic leadership shift at the White House signals an evolving regulatory environment that AI governance and compliance professionals will need to navigate.

3 · Skills, tools & techniques

Workplace AI agents are getting real access to tools, APIs, and infrastructure, so control matters. Three concrete skills to learn this week: (1) execution-time agent containment — sandbox the runtime, not just the prompt; (2) agent supply-chain security — marketplace review and automated scanning are not sufficient, as the 26,000-agent fake skill demonstrated; (3) treating every AI agent as an identity, with the same access controls and audit posture you would give a human employee.

Bottom line

Today's theme is control. Control the agent runtime via execution-time safety kernels, control the software supply chain so fake skills cannot reach 26,000 agents, understand the shifting policy environment, and control your own career risk by building skills AI cannot easily commoditize.

⚙️ Automated brief, human-reviewed. AI-researched from HackerNews, TechCrunch AI, VentureBeat AI, and The Decoder; written by Mistral Medium 3.5 running locally on a Mac cluster; narrated by VibeVoice 1.5B (Maya solo and Maya + Carter dialogue). Every claim links to its source — click through to read the original.