Technology & innovation
Anthropic unveiled Claude Mythos Preview last week, describing it as the most capable AI model ever built and simultaneously announcing it would not be made available to the public. The reason: it can autonomously find and exploit security vulnerabilities in ways that could be catastrophic in the wrong hands.
During testing, Mythos found zero-day vulnerabilities in every major operating system and web browser, discovered a 27-year-old flaw in OpenBSD that had survived decades of human review, and autonomously chained Linux kernel vulnerabilities together to gain complete machine control. Developing a full root exploit from a known vulnerability cost under $1,000 and took half a day.
In one unsettling test, a researcher asked the model to find a way to escape its sandbox and send a message if it succeeded. It did, then went further without being asked, posting exploit details to public-facing websites. The researcher found out when he received an unexpected email from the model while eating lunch in a park.
We find it alarming that the world looks on track to proceed rapidly to developing superhuman systems without stronger mechanisms in place."
Anthropic is making Mythos available only to roughly 40 vetted organizations through Project Glasswing, a cybersecurity defense initiative. Launch partners include Amazon, Apple, Google, Microsoft, CrowdStrike, JPMorgan Chase and the Linux Foundation.
Competitors are expected to reach similar capabilities within six to 18 months, which Anthropic says is precisely why defenders need a head start now.
Read more via Axios, Forbes, MSN
AI's impact on workers without college degrees has been largely overlooked in public debate, and that the risk goes beyond individual jobs, according to a new Brookings Institution report.
AI threatens to fracture entire career pathways that have historically allowed lower-wage workers to advance.
More than 15 million workers without college degrees are currently in jobs with high AI exposure. About 11 million of those hold "Gateway" jobs, the middle-rung roles that connect entry-level work to better-paying careers.
Gateway jobs most at risk are concentrated in clerical and administrative work, which is also disproportionately held by women. Customer service representatives, secretaries and accounting clerks are among the roles most likely to be disrupted.
Workers without degrees account for more than 62% of people in Gateway jobs across the workforce, making this group central to the broader employment pipeline. If those jobs disappear, the pipeline of experienced workers for higher-level roles disappears with them.
About a third of these workers have low adaptive capacity, meaning they will likely struggle most to adjust if displaced.
"AI is not just reshaping the software developers," said one report co-author. "This is coming to every community."
Read more via Fast Company, Brookings Institution
A three-year Wharton/GBK Collective study of business leaders at U.S. companies with revenues above $50 million finds a significant gap between how senior executives and middle managers experience AI, one that researchers say explains why less than 10% of companies are capturing meaningful AI value at scale despite heavy investment.
45% of executives report significantly positive ROI from AI investments. Among middle managers, that drops to 27%.
56% of executives believe their company is adopting AI faster than competitors. Only 28% of middle managers agree.
Nearly two-thirds of executives say they've become "much more positive" about generative AI over the past year. Only 39% of middle managers say the same.
The gap exists partly because executives use AI for strategic synthesis and decision support, where it performs well, while middle managers are responsible for deploying it in messy day-to-day operations with uneven teams, legacy workflows and zero tolerance for errors.
Middle managers are also already overloaded: McKinsey research shows they spend less than 30% of their time on people leadership, with nearly half consumed by administrative work. Handing them an AI mandate without reducing that load first is, as one executive coach put it, "asking them to build the plane while flying it."
Read more via Harvard Business Review
Robots powered by generative AI are showing up in hospitals, hotels, factories and restaurants, and the gap between these systems and the clunky service robots of a few years ago is significant. A Harvard Business Review analysis based on 18 months of field research across 14 organizations outlines what's working and what isn't.
BMW has been deploying a humanoid robot in its South Carolina assembly plant since 2024, using it to carry and place fragile sheet-metal parts. During an 11-month pilot it contributed to the production of roughly 30,000 vehicles. The company is now upgrading to a newer model.
A robot named Robin is providing emotional support in 30 pediatric units and nursing homes, moving autonomously between rooms, answering patient questions in plain language, and running games for children using spoken responses.
The technology works best when applied to repeatable tasks in industries with chronic labor shortages, such as hospital logistics, hotel check-in, and quick-service restaurants, where the operational handoff to a robot is clear.
Early deployments reveal persistent challenges: 71% of companies cite high upfront costs as a major barrier, and robots have struggled with accents, background noise and unexpected requests in some hotel settings, sometimes increasing rather than decreasing staff workload.
Robots learn from interactions, which means employees or customers can inadvertently, or deliberately, teach them bad behaviors. Without guardrails, those behaviors can spread quickly across entire robot fleets.
A McKinsey senior partner argues that most organizations are making the same two mistakes: treating all work the same when applying AI, and spreading investment too thin rather than concentrating it where it creates the most value. The result is a lot of activity with little impact.
Senthil Muthiah says companies tend to apply agentic AI across the board, where it moves quickly through structured, rules-based tasks but stalls in areas requiring human judgment. The fix isn't less AI, it's better targeting.
For every dollar spent on AI technology, organizations need to invest roughly two dollars in change management, capability building and adoption to actually realize the benefits, according to McKinsey research.
70% of human skills remain essential even in heavily AI-augmented environments, and the productivity upside, which McKinsey pegs at $4.4 trillion globally, only materializes for organizations that treat the people side of transformation as seriously as the technology side.
Perhaps the most candid finding: no function within most organizations currently owns the job of creating, managing, tuning and eventually retiring AI agents. "This will become a new organizational capability in the future," Muthiah said. "As of now, there is no clear view on who should own this."
Read more via UC Today
At a recent VentureBeat event, technology leaders from MassMutual and Mass General Brigham described how both organizations got out of pilot mode and into AI production.
MassMutual's results are concrete. Developer productivity increased 30%, IT help desk resolution times declined from 11 minutes to one, and customer service calls cut from 15 minutes to one or two. The key was refusing to move to production until metrics were defined and a business partner signed off.
MassMutual also deliberately avoided locking into any single AI model, building common service layers between the AI layer and underlying systems so newer models can be swapped in without starting over.
Mass General Brigham took the opposite approach at first, letting pilots proliferate without governance, then course-corrected by shutting most of them down. A pivotal realization: they were building in-house tools that their existing vendors, including Epic, Workday and Microsoft, were already planning to provide.
In clinical settings, the guardrails are absolute: AI never issues a final decision. A physician always closes the loop.
Read more via VentureBeat
A Clayton County prosecutor admitted to using AI to draft legal briefs that contained at least five fabricated case citations and several misrepresented ones. The district attorney formally apologized to the Georgia Supreme Court, and the prosecutor faces potential discipline and State Bar referral. Legal experts say the incident is less a fluke than a warning sign of what's coming as AI quietly spreads through the legal system.
About 60% of judges are believed to be using AI in some capacity, according to one legal tech expert, for tasks ranging from legal research to drafting and case analysis. "It's been a quiet, rolling thunder," she said.
High-volume court systems like those in metro Atlanta are particularly likely to accelerate AI adoption, which raises the stakes for errors going undetected.
The legal responsibility is clear: attorneys are accountable for everything they submit, AI-generated or not. The Georgia State Bar has issued guidance reminding lawyers that AI cannot replace their professional judgment and that they remain responsible for verifying all AI-generated work.
Supporters argue AI could dramatically lower legal costs and expand access to justice. Critics worry about bias embedded in training data, overreliance on automation, and the erosion of human judgment.
Read more via CBS Atlanta
Accountants still have to physically count inventory, and they're desperately hoping AI will eventually save them: Despite widespread AI adoption in accounting and auditing, there's still no technological substitute for sending junior auditors into grain bins, freezers, snake-infested quarries and chicken coops to manually count things. Auditors describe ending up covered in corn dust, manure and fertilizer, or shivering through frozen fish warehouses. "I think AI can count something faster than a human can, so you can see it coming," said KPMG's U.S. assurance leader. For now, though, it remains a rite of passage for young accountants. (The Wall Street Journal)
Manufacturers are using AI translation tools to communicate with workers who don't speak English: A growing number of U.S. manufacturers are deploying AI-driven translation technology to bridge language gaps on the production floor, including real-time captioning at company meetings, translated safety signage, and smart earpieces for supervisors. Manufacturers employ roughly 3.1 million foreign-born workers, representing about 20% of the industry, and businesses lose an estimated $500,000 per year on average in hidden labor costs tied to ad-hoc translation practices. "No one had asked them how they're doing, how they're feeling, what they need, because no one has the Arabic skills to talk to them," said one consultant. "Now they're being seen and heard." (HR Dive)
Utah is letting an AI chatbot renew certain psychiatric prescriptions: Under a new pilot agreement with Legion Health, Utah is allowing an AI system to authorize renewals for a limited set of psychiatric maintenance medications, including common antidepressants like fluoxetine and sertraline. The chatbot cannot diagnose, start new treatment, change doses, or switch medications, and any case involving suicidality, worsening symptoms or pregnancy must be escalated to a human provider. The pilot is heavily staged, requiring a licensed clinician to review each of the first 250 cases before anything goes to a pharmacy. Utah is framing the agreement as regulatory mitigation, not endorsement, and explicitly states it does not waive liability if a patient is harmed. (TechRepublic)
