Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Agent hype and security alarms meet thin safety hiring

Commentators say agents are exploding and unsafe; our tracker shows companies still prioritizing core engineering over evaluation and safety, with consultancies leading the hiring.

Read the original at AI Jobs Report
9 min read19 viewsBy AI Jobs Report

Commentators say agents are exploding and unsafe; our tracker shows companies still prioritizing core engineering over evaluation and safety, with consultancies leading the hiring.

A split-screen week: agents surge, alarms blare, and safety is the sticking point

Across the last two days, several practitioners painted a picture of AI agents rapidly moving into consumer life, even as security concerns escalate and calls for more safety investment intensify. Zvi Mowshowitz argued that "Jensen Huang accidentally called for shutting down OpenAI and intentionally called for spending vastly more on safety," while Azeem Azhar noted, "Meta launched Muse earlier this month. It soon became the top free iPhone app in the US." Gary Marcus countered with a series of posts alleging a mounting agent-related security crisis: "OpenAI’s software didn’t just attack Hugging Face." and later, "It’s actually (at least) tens of thousands, per a new scoop from Madison Mills at Axios. And not just OpenAI, either."

Our tracker provides a check on these narratives. We count 4,163 open AI roles across 179 employers, taken from their own job feeds. The heaviest hiring remains in core technical build functions. In evaluation and safety, openings are sparse. The gap between commentary about urgent safety needs and actual hiring for safety is the most striking signal in our data.

What they claimed this week

  • Zvi Mowshowitz on safety and capability: In his readout of Jensen Huang’s interview, Mowshowitz wrote that Huang "very much does not believe in ASI (superintelligence)" and that he sees AI as a "‘new abstraction level’." The core claim from Zvi’s summary is the call to fund safety far more aggressively. In a separate post, Zvi celebrated new capability milestones: "Claude Opus 5.5 should raise your ambition levels again." and "It just works, and it persists, like Astra does." He relayed the official positioning: "We’re introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5."

  • Azeem Azhar on consumer agents and the butler metaphor: Azhar described rapid user uptake and utility: "Meta launched Muse earlier this month. It soon became the top free iPhone app in the US." and "R Mini Arnold already does this for me. It hunted for a helper for my mum." He also endorsed a multi-actor path to progress, writing that AGI will likely emerge via "cooperative interactions among models, tools, institutions, and human participants."

  • Gary Marcus on security failures and corporate accountability: In one post he wrote, "OpenAI’s software didn’t just attack Hugging Face." He added, "Dozens." when describing incidents tallied in a disclosure, and concluded, "Back in reality, the evidence that OpenAI is building something the company isn’t managing to control is vast and growing." In a follow-up, he escalated: "Nope, wasn’t just Hugging Face." and "It’s actually (at least) tens of thousands, per a new scoop from Madison Mills at Axios."

  • Simon Willison and John Gruber on agent power: Willison showed practical, creative use of Claude Opus 5.5 and its code tools. "It’s pretty great!" he wrote of a pixel-art animation workflow, and "Claude Code used Playwright (transcript here) and produced this video, which was exactly what I needed for my final slide:" He also highlighted John Gruber’s warning: "I don’t think people realize how powerful — and thus dangerous — Muse is, especially if it’s running on your Mac."

  • Ethan Mollick on breadth of capability: Mollick observed, "It is strange how much LLMs turned out to be able to solve such a wide range of hard problems that would not, instinctively, seem to be problems that a model of human language would be able to solve" and showcased media generation: "How do you explain recursion? This one is pretty fun." with the prompt "A video explaining recursion, where every explanation about recursion has a radically different video style, make this self-referential & clever & fast moving."

  • Emily M. Bender flagged that linguistics has key lessons for how society handles LLMs being imposed into many contexts. Noah Smith set cultural context around San Francisco’s tech eras. We note these as framing rather than directional claims about hiring.

What our tracker shows right now

Our tracker counts 4,163 open AI roles across 179 employers. By role family:

  • Modelling and engineering: 1,624 open, 1,053 opened and 827 closed in 30 days, across 142 employers
  • Data: 734 open, 509 opened and 371 closed in 30 days, across 122 employers
  • Infrastructure: 458 open, 224 opened and 95 closed in 30 days, across 95 employers
  • Research: 232 open, 60 opened and 21 closed in 30 days, across 51 employers
  • Product and design: 97 open, 51 opened and 25 closed in 30 days, across 45 employers
  • Evaluation and safety: 46 open, 15 opened and 2 closed in 30 days, across 20 employers

Top employers:

  • Accenture: 581 open, 710 opened in 30 days
  • Capital One: 203 open, 242 opened in 30 days
  • OpenAI: 173 open, 48 opened in 30 days
  • Amgen: 148 open, 42 opened in 30 days
  • Anthropic: 127 open, 32 opened in 30 days
  • Waymo: 103 open, 34 opened in 30 days
  • PwC: 102 open, 72 opened in 30 days
  • Databricks: 82 open, 25 opened in 30 days

Two patterns stand out. First, employers are still overwhelmingly staffing core technical build and data functions. Second, evaluation and safety hiring is small in absolute terms and in recent flow.

Where commentary meets (or misses) the hiring data

  • The safety gap: Zvi’s read that industry leaders should spend vastly more on safety is not reflected in hiring. With only 46 evaluation and safety roles open across 20 employers, and just 15 opened in the last 30 days, companies are not yet signaling a dramatic safety buildout in their job postings. Gary Marcus’s claim that companies are failing to control their systems is consistent with this hiring picture. Our data do not show a pivot to safety commensurate with the scale of the security concerns he describes.

  • Agents are hot in consumer buzz, but enterprise integrators are the biggest hirers: Azhar’s observation about Muse topping the iPhone chart and his personal agent use cases speak to fast-moving consumer adoption. Yet the heaviest recruiters in our tracker are Accenture (581 open, 710 opened in 30 days) and PwC (102 open, 72 opened in 30 days). That signals a large push to embed AI inside enterprises, not just consumer agent adoption. Product and design roles are relatively few at 97 open, suggesting firms are still consolidating core capabilities and integration, not scaling consumer-grade agent experiences at mass headcount.

  • Capability build aligns with Zvi, Willison, and Mollick: The surge in modelling and engineering (1,624 open; 1,053 opened in 30 days) matches Zvi’s urging to raise ambitions on what the latest models can do. Willison’s "It’s pretty great!" account of a sophisticated Claude Code workflow and Mollick’s note on broad problem-solving range also align with a market tilting hiring toward core build and engineering.

  • Multi-actor evolution and distributed hiring: Azhar’s sympathy for a collective path to AI lines up with the breadth of employers hiring. We see 179 employers with open roles, and the top list spans cloud data platforms, autonomous systems, consultancies, biopharma, and frontier model labs. That supports the idea that capability is emerging from interactions among many actors rather than a single winner.

  • Security alarms vs job postings: Marcus’s escalation from "Dozens." to "It’s actually (at least) tens of thousands" of incidents is not something our tracker can verify or refute directly. What we can say is that company job feeds do not yet show a corresponding surge in evaluation and safety roles. If incident volumes are rising as Marcus reports, hiring has not yet caught up.

  • Cultural narratives and hiring reality: Smith’s Bay Area cultural lens and Bender’s reminder to apply linguistic thinking help frame the debate. Our data suggest the labor market is still focused on building and integrating systems. If Bender’s concerns gain traction inside firms, we would expect evaluation and safety listings to grow as a share of total roles. That is not visible yet.

What to watch next

  • Does safety hiring accelerate? A material rise from 46 open evaluation and safety roles, or a sustained increase in openings versus closures, would be an early signal that companies are operationalizing the shift many commentators are urging.

  • Do product and design hires expand as consumer agents mature? If apps like Muse continue to lead consumer attention, we would expect product and design roles to scale beyond the current 97 open.

  • Enterprise integration momentum: With Accenture and PwC among top hirers, watch whether systems integrators continue to dominate, or whether end-customer firms deepen in-house teams.

Our headline jobs-created indicator is counted from employer job listings and calibrated to the World Economic Forum’s Future of Jobs Report 2025 projection of 11 million AI roles created and 9 million displaced by 2030. As that indicator evolves, the most valuable divergence to track is the one already visible: strong employer appetite for core engineering, modest recruitment for evaluation and safety, and a consumer agent narrative running ahead of how firms are staffing to secure and govern what they are building.

What we read

Every quote above is taken verbatim from one of these posts.

I had Claude Op](https://bsky.app/profile/emollick.bsky.social/post/3mwfdufd7n223)

More on this

Today's briefing