WORKFORCE

Can we preempt an AI-induced economic meltdown?

As concerns continue to mount about the impacts that AI could have on the economy, OpenAI's philanthropy arm wants to front-run the crisis. 

On Wednesday, the OpenAI Foundation announced a $250 million commitment to creating grants and forging partnerships aimed at helping prepare the economy for AI disruption. The goal, the foundation said in its announcement, is “building secure and abundant economic futures.” 

The initial investment will support external organizations through grants, open calls, and institutional partnerships. The foundation will also build an internal team to advance this work directly, and expects the first initiatives to be announced later this year. 

“AI is going to lead to huge economic changes as it makes previously scarce capabilities far more widely available, and there is deep uncertainty about how far and how fast they will go,” OpenAI said in its release. “We don’t need to know exactly how the future will unfold to prepare for it.” 

The program, which aims to create “institutional options” for AI preparedness, will work across three main areas: 

• Understanding the shift: Here, the foundation will invest in independent measurement and forecasting to develop a clear understanding of how this tech will impact the economy. This involves measuring how people leverage AI and what it enables them to access, rather than just the earnings AI can generate. 

• Supporting the transition: This involves offering resources to workers as they navigate near-term disruption. Along with support and retraining, OpenAI notes that this includes researching approaches that give workers “agency over AI deployment,” and investing in broadening the capacity of governments and public institutions to deliver. 

• Building for long-term economic security: This part involves developing approaches to organizing “post-AI political economies” and determining how to equitably share global economic gains. This includes researching proposals to shift taxes and dividends in response to “observable indicators,” and figuring out ways to give people “durable claims” on broad economic growth. In other words, it's talking about how to share the wealth of AI. See similar proposals made in OpenAI's 'AI New Deal' document.

With studies and forecasts regularly published about the tech’s ability to upend the job market, the growing fear of AI’s looming impacts on the economy is breeding distaste and resentment towards the technology. The foundation’s end goal is to strengthen the global economy and the workforce’s ability to handle change by understanding which approaches work before damage is done and a crisis is precipitated.

The OpenAI Foundation is trying to get ahead of problems that tech from OpenAI and other frontier labs could accelerate in the years ahead, if the current trajectory plays out. And, for multiple reasons, it’s in OpenAI’s best interest to untangle these issues. For one, with anxiety and animosity towards AI worsening, a move like this is good PR, seeking to ease those fears by positioning OpenAI on the side of the people. Additionally, what goes around comes around: a weakening economy is bad for everyone and would hinder OpenAI’s ability to sell its services. And broadly, while OpenAI doesn't have the same reputation for responsibility as rival Anthropic, its stated mission has always been to create AGI that benefits all of humanity. Given the forecasts that this tech could widen economic disparity, getting in front of that will be vital to delivering on that mission. 

TOGETHER WITH PLAID

3 ways to stay ahead of next-gen financial fraud

Generative AI is helping fraud evolve faster than traditional defenses can handle. But these same tools can be put to good use. Read MIT Technology Review and Plaid’s report to see how technology is reshaping financial defenses. 

Read the report to discover:

• AI-powered fraud tactics targeting your customers

• Why layering AI into your defense is key

• How industry collaboration is reshaping fraud prevention

Get the report

RESEARCH

New Cisco research challenges AI safety benchmark

Nearly all new AI models are released alongside model cards that feature safety benchmarks, but new research calls into question at least one of those measurements, claiming it targets the wrong criteria.

Cisco research released Wednesday concludes that no frontier AI model is iteratively safe, a finding drawn from an evaluation of how large language models from leading labs respond to both single-turn attacks, commonly used in benchmarking, and iterative, multi-turn attacks, which more closely resemble real-world adversarial behavior.

The results made it clear: the single-turn attack success rate (ASR) is not a reliable proxy for what happens in the real world when an attacker iterates in real time. 

“Single-turn ASR has been the default because it is a simple and easily reproducible metric that matched early prompt injection and jailbreak threat models,” Amy Chang, head of AI threat and security research at Cisco, told The Deep View. “While still a useful metric, it is no longer adequate on its own — as these considerations break down in a multi-turn scenario — and single-turn ASR does not serve as a proxy for a model’s multi-turn resilience.” 

The results draw from a paired-regime evaluation of 15 closed/proprietary models from OpenAI, Anthropic, Google, Amazon, and xAI. Each model was exposed to 30,090 single-turn prompts (2,006 per model) and 6,986 multi-turn attacks distributed across 1,456 conversations, using a shared harness, prompt bank, and the Cisco Integrated AI Security and Safety Framework taxonomy.

The findings were consistent across all models tested: multi-turn ASR ranged from 7.89% to 88.30% (lower is better) across the cohort, while single-turn ASR ranged from 2.19% to 64.91%. Some standalone model performance highlights included: 

• Amazon Nova 2 Lite: Had the lowest multi-turn ASR at 7.89% 

• Anthropic Claude family: Despite having the lowest single-turn refusal (2.19% to 3.64% ASR), it reached 11.16% to 16.20% with multi-turn attacks 

• OpenAI GPT-5.4: This model showcased a 9x increase in ASR with multi-turn attacks, moving from 2.74% single-turn to 24.68% multi-turn

While some results sit at the lower end of the spectrum, Amazon Nova 2 Lite's performance being a notable example, they still represent meaningful residual risk, reinforcing the report's central conclusion that no model is inherently safe. This finding also aligns with Cisco's recent research. A multi-turn red-teaming study found that vulnerability rates rose 71% after five-turn conversations compared with single-turn evaluations.

The call to action for users: be as aware as possible of potential hidden risks and take adequate precautions. 

“No base model is iteratively safe, which means defense-in-depth is the price of deploying AI securely,” added Chang. “Depending on your organization’s use case and AI strategy, this may mean: the use of runtime guardrails; additional input/output monitoring; red-teaming models, applications, and agents; and application-layer policies.”

The broader takeaway of this paper is one I keep returning to in this space: current benchmarks test only narrow, highly specific tasks that don't reflect how models are actually used in the real world. The implications are significant, ranging from overstating a model's intelligence (a model that excels at physics problems, for instance, may struggle with something as basic as a natural conversation) to creating genuine security risks, as outlined above. This isn't an argument against benchmarks, as they remain valuable. However, it does expose a gap in the industry: the need for more standardized, representative evaluation frameworks. Whether that comes through regulation or expanded third-party testing, the status quo isn't enough. This area needs to expand and focus on more real-world scenarios.

IN PARTNERSHIP WITH LAMBDA

Most AI teams treat compute as a commodity. It’s not.

Compute looks interchangeable on a spec sheet. At scale, the facility, the cluster design, and the engineering behind it determine whether you ship on schedule or burn budget.

The teams pulling ahead know that compute infrastructure shapes what you can build, how fast, and at what margin. Lambda was built by researchers who learned this firsthand.

Find out why.

STARTUPS

Robot makers just got a stronger AI stack

With AI companies lining up to claim a slice of the enterprise market, Mistral has set its sights on a new frontier: robotics, physical AI, and industrial engineering.

On Thursday, the French AI lab launched Mistral for Industrial Engineering, a fully integrated AI stack that combines advanced models, engineering expertise and robotics to assist with industrial operations. Ultimately, this offering aims to help engineers customize frontier models on their data and assets, such as drawings and blueprint files, and use physics-aware synthetic simulation models. 

Arthur Mensch, co-founder and CEO of Mistral, claimed in a blog post that industrial engineering is the “heart of the next AI revolution,” noting that Mistral’s new product allows firms to get the most out of robotics and physical AI by customizing it with their own data and deploying it within their own infrastructure. 

"With Mistral for industrial engineering, we put AI at the center of the physical product engineering lifecycle,” Mensch said. 

This launch was made possible in part by Mistral’s recent acquisition of Emmi AI, an Austrian AI startup with expertise in physical AI and state-of-the-art engineering models. Some of the practical use cases include assisting with design, production, quality inspection, and validation. It also supports agentic workflows tailored to mission-critical engineering environments. 

Mistral hosts the infrastructure itself on private bare-metal servers, bundling GPU capacity, reference architectures, and tested operating patterns, according to the blog post. This direct connection to customer networks keeps sensitive data under control while supporting hybrid setups as workloads scale.

Alongside the announcement, Mistral announced partnerships with Airbus, which will implement Mistral’s AI at the core of its operations and processes, and with BMW Group, which will use Mistral as a central partner for its “Large Industry Model” initiative.

This isn't Mistral's first industry-specific push, having launched Mistral for Finance earlier this year. The strategy makes sense for most AI labs: vertical offerings target a defined audience, making it easier to demonstrate direct value and address specific needs rather than competing for general consumer attention. That's particularly true for Mistral, which has always focused on enterprises rather than chasing the broader consumer market, giving it a competitive edge. Its European roots give it another advantage, too, as stricter data sovereignty requirements make it a more credible choice for enterprises with sensitive data, including those in the US and other countries.

LINKS

• Robinhood users can connect agents to make credit card purchases

• YouTube makes AI-generated content labels more prominent 

• Meta laid off 1,400 Washington employees in the wake of AI restructuring 

• Huang says Nvidia to spend $150 billion a year in Taiwan 

• AWS sells AI shopping technology to other retailers 

• AI defense startup Airis Labs raises $60 million, $31M million Series B

• Career expert: These five skills are AI-proof and getting more valuable

• DuckDuckGo installs jump 30% as users balk at Google’s AI search overhaul

• The job AI is actually increasing: cybersecurity experts

• AI psychosis may be hitting Tech CEOs the hardest

• YouTube will start automatically labeling AI-generated videos 

• Lovable: Launches subagents to run projects in parallel 

• OpenClaw 2026.5.26: Updates include lower-latency replies, more reliable channel flows

• Krea 2: The API is available on fal, ComfyUI, and Nous Research  

• ElevenLabs: Speech Engine Skill can give agent a voice without leaving your LLM stack

• NotebookLM: Now syncs to Google Drive

• Sesame Personal Agents: These agents are built to be more natural and conversational

• Cisco: AI Security Engineer

• University of Chicago: AI Application Developer

• Lilly: Sr. Staff Software Engineer - AI Chat

• CoreWeave: Staff AI Security Engineer

GAMES

POLL RESULTS

Would you see an AI-generated film in a theater?

Yes (23%)
Maybe (30%)
No (40%)
Other (7%)