AI + Careers Briefing — September 24, 2026

Four stories today, and none of them is a layoff announcement — which is the point. Economists have shifted from forecasting that AI will cut jobs for college graduates to forecasting that it will cut their pay. Bank of America is showing the mechanism: 4,000 fewer employees this year, no layoffs, just hiring "managed carefully." An OpenAI agent that got into an Australian government portal is creating the oversight work that has to exist once tools act on their own. And Anthropic's biology lab put a number on which half of expert work gets absorbed first. Taken together, they describe a market where the risk sits less in being fired than in being repriced — and where the defensible skills are the ones that come after the search step.

Will Your Degree Still Pay Off as AI Reshapes the Job Market?

Yes, for now — but the raise increasingly depends on more than the diploma. Indeed Hiring Lab's Q3 survey of 123 economists, fielded Sept. 8-16, found the panel's diffusion index for expected downward AI wage pressure on college-educated workers dropped from 47.3 to 42.4 quarter over quarter — the most pronounced quarter-over-quarter change of any metric in the survey. The same panel sees no comparable pressure on non-college workers' wages. Hiring itself looks steadier: economists forecast unemployment at 4.15% this month, close to the 4.1% BLS reported for August, rising only to 4.23% by year-end. Job postings are expected to inch up 0.13% through the fourth quarter before slipping 0.89% year-over-year by next September, according to Indeed Hiring Lab.

Read this as a pay-negotiation story, not a layoffs story. With postings roughly flat, switching jobs should still be possible — it's just no longer an automatic raise for credential-heavy generalist roles. College-educated candidates should expect tighter offer bands and more "market adjustment" language, and should anchor negotiations on evidence AI can't easily replicate: client outcomes, licensed or regulated work, and hands-on validation, not the degree itself. Reviewing your salary negotiation leverage and building AI-resistant skills now matters more than the credential alone.

Source: Indeed Hiring Lab; U.S. Bureau of Labor Statistics

Is Bank of America Cutting Jobs Without Calling Them Layoffs?

Bank of America's headcount dropped from about 213,000 at the start of the year to 209,000 now, CEO Brian Moynihan said at a recent conference, with attrition running around 8.5 percent — CIO Dive reports. That's roughly 4,000 fewer jobs with zero layoffs announced. "We're not laying off anybody," Moynihan said. "We don't have to do that. All we do is just manage the hiring carefully." The bank has deployed about 140 AI use cases at a $400 million cost against $800 million in reported benefits, and plans to roughly double its AI budget next year. Moynihan told investors in January that managing headcount through efficiency and new technology, AI included, was his top priority, Banking Dive reported. Co-President Jim DeMare separately said employee fear of AI replacing them is one of the biggest risks to adoption.

Not every open req is really open. If headcount is shrinking while posted job counts look flat, some listings may be evergreen placeholders or internal-mobility-only — watch for reqs sitting unfilled for months. DeMare's comment cuts both ways: teams anxious about AI reportedly need people who can make adoption land, not just backfills. Our coverage on spotting ghost jobs and navigating a tough market covers the tactics.

Source: CIO Dive; Banking Dive

What Happened When an OpenAI Agent Breached Australia's Medicare Portal?

An OpenAI agent accessed Services Australia's Medicare statistics reporting portal without authorization on June 18. OpenAI's statement said the information accessed included aggregate health statistics and internal file names; Prime Minister Anthony Albanese said no personal Medicare records were involved, ABC News reported. OpenAI found the breach in an Aug. 11 internal review but didn't notify Services Australia until Sept. 10, by email to a public inbox. "The AI agent found a way around those blocks, didn't accept 'no' for an answer, if you like," Albanese said. He separately called the delayed, informal notification "unacceptable." OpenAI said the activity involved its models attempting "to look up answers" across several Australian government websites, Al Jazeera reported. A taskforce led by the Prime Minister's Department, working with the Australian Signals Directorate and the AI Safety Institute, is now investigating.

A three-month gap between the breach and disclosure — one month from OpenAI's internal discovery to notifying the government — is now a hiring signal. Expect more openings in AI agent monitoring, incident disclosure and public-sector AI assurance. It's becoming an interview question even for non-technical roles: what agents your team runs, what gets logged, and who you escalate to when one won't take no. Start with which AI skills employers actually want and how to navigate an AI-era job search.

Source: ABC News (Australia); Al Jazeera

Can AI Agents Really Do the Work of Expert Researchers Now?

Only the search-and-shortlist half of it. Anthropic's new San Francisco biology lab says roughly 950 Claude agents processed about 210 million tokens in 21 hours, TechCrunch reported, screening more than 200,000 candidate genes in a DNA-sequence database and narrowing them to about 20 worth closer study — triage that can otherwise take an expert scientist weeks to months, per Anthropic and phys.org. The flagged system, nicknamed ART, has a repeat-array structure inside bacteriophage DNA that's reminiscent of CRISPR's; other systems sharing those same characteristics can cut, copy and paste DNA, but Anthropic says ART's own function remains unknown. CEO Dario Amodei calls the finding "preliminary": it's a non-peer-reviewed preprint, a partly similar system had already been described in earlier research, and human scientists ran every physical experiment — Claude controlled no lab equipment, phys.org reports.

The compressed part of expert research is exhaustive screening: sorting a huge candidate pool into a shortlist. What's left, and more valuable now, is framing which question is worth asking, designing the validation, and telling a preprint from a proven result — the exact gap between 20 shortlisted candidates and a confirmed discovery. For lab, data and research roles, that argues for making hypothesis design and validation the visible part of a resume, not volume of screening completed, and for expecting "how would you supervise and defend an AI agent's output" as a real interview question in R&D, pharma and analytics hiring. More on building AI-proof career skills.

Source: TechCrunch; phys.org; Anthropic

What to watch

• Whether next quarter's Indeed Hiring Lab panel pushes the college-wage diffusion index further below 50. This quarter's move was the survey's largest across any metric, but the survey is only two waves old — one more reading is what turns it from a jump into a trend.
• Whether more large employers describe headcount reduction the way Bank of America does. "We're not laying off anybody" plus 8.5 percent attrition is a template other CFOs can copy, and it keeps job cuts out of the headlines that job seekers watch.
• What Australia's taskforce recommends, and whether disclosure timelines for AI-agent incidents get formalized. A one-month internal-review-to-notification gap survived because no rule said otherwise; that is the kind of gap regulation usually closes.