AI Watch
News Source
EXCERPT:
What began with GPUs has expanded into full-stack AI factories comprising accelerated compute, high-speed interconnects, liquid-cooled systems, inference software, autonomous agents, reference architectures and the ecosystem needed to build and operate them at scale.
Full-stack AI factories are part of the broader ecosystem that NVIDIA is helping define and build. NVIDIA closely collaborates with global system partners such as Cisco, Dell, HPE, Lenovo and Supermicro to bring AI infrastructure to enterprise data centers. NVIDIA also relies on a curated ecosystem of AI software partners to build AI solutions for each enterprise’s use cases. This ecosystem supports a choice of models, across proprietary and open options.
News Source
EXCERPT:
Artificial Analysis and IBM Software Innovation Lab are launching ITBench-AA, the first in a new series of benchmarks evaluating models on agentic enterprise IT tasks, starting with Site Reliability Engineering tasks where frontier models score below 50%
ITBench-AA’s SRE tasks benchmark model performance on Kubernetes incident response, where models and agents must diagnose live systems by reading logs, tracing dependencies, and identifying root-cause entities across complex infrastructure. The underlying ITBench dataset has been developed by IBM, leveraging deep expertise in enterprise IT operations.
Two major AI security flaws have been revealed; corporations aren’t building safeguards into their AIs and there is a security vulnerability in the code itself.
News Source
EXCERPT:
China now requires people working in AI at private firms to secure travel approval before leaving the country. According to Bloomberg, the restrictions apply to individuals working in state-owned firms, startup founders, and those employed by private companies, as the central government considers them important strategic assets. China has already been limiting international travel for key individuals such as senior researchers at public educational institutions, nuclear scientists, and even top executives of government-owned companies, but extending the restriction to private firms and individuals is an uncommon move, even for Beijing.
There’s no official guidance yet on which roles, expertise, or seniority will be included in the travel ban. However, Bloomberg sources say that the individuals added to the list were assessed based on their impact on China’s AI ambitions, not just where they work or their position within their company. This move is an expansion of a former government directive wherein some AI engineers had mandatory reporting of any overseas travel plan, although they were still free to go abroad as needed.
This shows that Beijing considers AI as a strategic advantage and that the people leading the industry are considered crucial for the country’s advancement. This news comes months after Meta’s surprise purchase of Manus AI, which China wants to unwind to prevent the U.S. from acquiring Chinese AI talent and intellectual property. Although the two aren’t directly related, the report says that the new policy is designed to protect against the leaking of key technologies, such as the one being developed by the Chinese startup that moved to Singapore.
News Source
EXCERPT:
Millions of AI agents and tools around the world have been imperiled by a critical vulnerability that can allow hackers to breach the servers running them and make off with sensitive data and credentials to third-party accounts, a security researcher is warning.
The vulnerability is present in Starlette, an open source framework that its developer says receives 325 million downloads per week. Thousands of other open source projects are also vulnerable because they require Starlette to work. The framework is an implementation of the ASGI (asynchronous server gateway interface), which allows large numbers of requests to be efficiently processed simultaneously. Starlette is the base of FastAPI and other widely used frameworks for building services in Python apps, as well as many others.
Trivial to exploit, millions of servers exposed
ASGI, and by extension Starlette, have access to servers running the MCP (model context protocol), which allows AI agents from major providers to access external sources, including user data bases, email and calendar accounts, and all manner of other resources. To connect with these external systems, MCP servers store credentials for each one, making them especially valuable storehouses for attackers to breach.
The vulnerability, tracked as CVE-2026-48710 and under the name BadHost, is trivial to exploit and works against most systems that aren’t behind a properly configured firewall. Besides FastAPI, other widely used packages—including vLLM, and LiteLLM—are also affected. BadHost affects Starlette versions prior to 1.0.1, which was released Friday.
“A single character injected into the HTTP Host header bypasses path-based authorization in Starlette, the routing core of FastAPI,” researchers from Secwest wrote. “Through FastAPI, this primitive (now tracked as CVE-2026-48710 and branded BadHost by the discoverers) reaches a large segment of the Python AI tooling ecosystem: vLLM (where the bug was discovered), LiteLLM, Text Generation Inference, most OpenAI-shim proxies, MCP servers, agent harnesses, eval dashboards, and model-management UIs.”
BadHost carries a severity rating of 7 out of 10. Secwest said the classification “materially understates” the threat it poses to people using other apps that depend on Starlette. X41 D-Sec, the security firm that discovered it, described it as having “critical severity.” X41 D-Sec partnered with fellow security firm Nemesis to create an online scanner that can check if a given server is vulnerable.
News Source
EXCERPT:
Chompie, one of the world’s tops ethical hackers, says AI like Claude Mythos will make it harder for people like her to compete.
News Source
EXCERPT:
In late March around 15 religious thinkers met with the artificial intelligence company Anthropic to discuss one of the strangest and most consequential questions now facing the AI industry: How do you teach a chatbot to be good?
The invitations to these meetings had arrived in different ways. Greg Cootsona’s came via e-mail. Brian Patrick Green’s came via a friend of a friend after Anthropic asked for suggested names. Both ended up in a series of conversations with the company about Claude, Anthropic’s chatbot, and the moral framework meant to guide how it behaves.
The aim wasn’t to make the chatbot Bible-thumping or pious. But it was an acknowledgment that centuries-old traditions of moral reasoning might offer insights to a five-year-old frontier AI lab whose systems are becoming more capable, more persuasive and harder to govern by simple rules.
News Source
EXCERPT:
Abstract: Subject-driven image generation aims to synthesize new images that preserve the identity of the given subject while following textual instructions. Existing approaches often encode text and reference images separately. This limits cross-modal reasoning abilities and causes copy-paste artifacts. Recent frameworks that connect multimodal models and diffusion models improve instruction following, but largely overlook identity preservation. To address these limitations, we condition diffusion models on Multimodal Large Language Models (MLLMs) that jointly encode text and reference images, and augment it with VAE-based identity conditioning. A novel Dual Layer Aggregation (DLA) module is designed to aggregate multi-level MLLM features for optimal conditioning, and a multi-stage denoising strategy is applied to progressively balance the semantic information from MLLM and fine-detail identity from VAE during inference.
News Source
EXCERPT:
As AI agents are integrated into an organization, enterprises will need to pivot from a set of linear processes and steps, to rewiring work in a very different way, explains Shah. That’s because the value in AI agents isn’t as another layer in an existing technology stack but as a connective tissue, he explains, moving between or across layers to coordinate a high-level task or retrieve and interpret data from multiple discrete applications. AI agents can create “a true competitive differentiation for an enterprise” by making decisions based on this capacity to contextualize, he says. “That is where the next battleground will be.”
To build this connective tissue, leaders need to adapt their technology stack to surface higher quality decisions from AI agents, prioritizing access to multiple datasets and applications simultaneously to develop tacit knowledge. “Organizations that make this architectural shift become genuinely more adaptive,” says Chatterjee. “When a new business requirement emerges, you don’t wait six months for a software vendor to build a feature. You configure an AI employee using natural language and connect it to the systems it needs. The time from business to production workflow drops from months to days.”
News Source
EXCERPT:
Pope Leo XIV called Monday for robust regulation of artificial intelligence and for its developers to work for the common good rather than profit, issuing a sweeping manifesto on safeguarding humankind as the technology impacts everything from work to war.
“Magnifica Humanitas” (Magnificent Humanity), Leo’s first encyclical, has been eagerly awaited ever since history’s first U.S.-born pope announced days after his election that he considered AI to be the biggest challenge facing humanity today.
In the text, Leo denounced the “culture of power” driving the AI race, especially in developing ever more sophisticated methods of remote warfare. He declared that it was “not permissible” to entrust irreversible, lethal decisions to AI systems, setting up another flash point between the American pope and the Trump administration, which has worked aggressively to deregulate AI development.
“Artificial Intelligence now demands to be disarmed, freed from logics that turn it into an instrument of domination, exclusion and death,″ the pope told a special Vatican presentation of the encyclical, one of the most authoritative types of teaching documents a pope can issue.
Experts in the tech industry, academia and Catholic morality said the document will likely become a benchmark in the debate over AI, a point of reference for policymakers, researchers and ordinary folk alike. It comes as the near-daily developments in the technology trigger concerns over AI replacing human jobs and even human intelligence.
Taylor Black, a Microsoft AI executive and director of Catholic University of America’s AI institute, said the document would prompt people “at the forefront of these tools” to ask questions such as “What does it mean to be human?”
News Source
EXCERPT:
There is a category of production incident that engineering teams are not tracking yet — because it doesn’t fit any existing postmortem template.
The agent initiated an action. The action was technically correct given the agent’s context. The context was incomplete. The infrastructure cascaded. And, by the time the incident review happened, three teams were arguing about whether it was an agent failure or an infrastructure failure, because the frameworks for thinking about these two things have never been connected.
The scale of this exposure is no longer theoretical. Seventy-nine percent of organizations now have some form of AI agent in production, with 96% planning expansion. Gartner predicts 33% of enterprise software will include agentic AI by 2028, but separately warns that 40% of those projects will be canceled due to poor risk controls.
What neither statistic captures is the failure mode happening between those two numbers: Agents that are running, that are not canceled, and that are quietly generating infrastructure events no one has categorized as risk.
News Source
EXCERPT:
Chinese AI startup DeepSeek just made one of the boldest pricing moves in the artificial intelligence race so far. The company announced it is permanently slashing the cost of its flagship V4-Pro AI model by 75%, bringing prices down to just a fraction of what developers were paying only weeks ago. AI companies worldwide have been facing two major problems: high infrastructure costs and limited access to high-end AI chips. So when a company suddenly cuts prices this aggressively — and permanently — it usually signals something important is changing behind the scenes.
DeepSeek says usage costs for V4-Pro now range from 0.025 to 6 yuan per million tokens, depending on workload type, down sharply from the previous pricing range of 0.1 to 24 yuan per million tokens. For developers building AI apps, agents, and services, that kind of drop could significantly lower operating costs.
News Source
EXCERPT:
Microsoft AI chief executive Mustafa Suleyman is warning that artificial intelligence could soon replace large portions of the white-collar workforce, predicting that AI systems will reach human-level performance across most professional tasks within the next 18 months.
The comments mark one of the clearest timelines yet from a major tech executive about how quickly AI could disrupt office-based professions, including law, accounting, marketing, and project management.
Speaking with the Financial Times, Suleyman said that most work involving “sitting down at a computer” is now vulnerable to automation as AI capabilities rapidly advance.
News Source
EXCERPT:
Pope Leo XIV has delivered a stark warning on artificial intelligence, claiming that the technology is aiding the “normalization of war” and transferring powers of life and death to unaccountable “technological actors.”
The American-born pontiff presented his warning on Monday in an encyclical titled ‘Magnifica Humanitas’ (Magnificent Humanity). In the 42,000-word document, Leo highlighted how the “growth of the military-industrial complex has become a defining feature of the current political landscape,” leading to “a troubling revival of war as an instrument of international politics.”
In this environment, “the development and use of AI in warfare must be subject to the most rigorous ethical constraints, to guarantee respect for human dignity and the sanctity of life and to avoid a race to develop such arms,” he continued.
News Source
EXCERPT:
A familiar warning now shapes much of the discussion about artificial intelligence: A handful of dominant firms will control the technologies, stifle innovation, and require aggressive antitrust intervention. It is a compelling story—and mostly wrong.
The idea that large companies automatically mean less innovation has become conventional wisdom in antitrust circles. European regulators have embraced it, blocking mergers and attacking American tech companies. The Biden administration followed that path, treating size itself as a threat and wanting government-led AI. The Trump administration, by contrast, has signaled a more evidence-based view—one grounded in both economic logic and empirical studies.
News Source
EXCERPT:
OpenAI CEO Sam Altman said on Tuesday the rapid development and adoption of AI would not lead to a global “jobs apocalypse” and the technology had not claimed as many white-collar jobs as he had feared.
Speaking virtually at a Commonwealth Bank of Australia (CBA) conference in Sydney, Altman said he was initially concerned about the impact AI would have on global employment levels.
News Source
EXCERPT:
Beginning this summer, University of California Berkeley School of Law students will be banned from using artificial intelligence to complete coursework or exams.
Under the newly adopted policy, students cannot “conceptualize, outline, draft, revise, and edit their work” using AI.
It also explicitly forbids students from asking AI to correct grammar mistakes or translate a paper into English.
Students are permitted to use AI for “research on papers ONLY for the limited purpose of identifying sources, such as cases, statutes, or secondary sources,” the policy states.
However, professors are permitted to make exceptions to this rule as long as they “do so in writing and with appropriate notice and require students to disclose any authorized AI use,” it states.
News Source
EXCERPT:
A proposed data center expected to cost more than $5 billion ran into intense resistance Thursday night in rural Pennsylvania, where residents packed a town hall meeting and delivered a clear message: They do not want their farmland and community identity sacrificed for a largely undefined mega-project.
During a three-hour informational session at Bangor Area Middle School, residents of Lower Mount Bethel Township voiced overwhelming opposition to the proposed Lower Mount Bethel Tech Center, according to WFMZ-TV. The event was organized by the project’s major stakeholders, including Peron Development and J.G. Petrucci Co., rather than township leaders.
News Source
EXCERPT:
China has launched a national programme that will assign every humanoid robot manufactured in the country a unique digital identity code, effectively a citizen ID, but for bipedal machines (those that can balance and walk/run on two legs).
The initiative, called the Humanoid Full Lifecycle Management Service Platform, was announced on Friday. It is led by the Humanoid Robotics and Embodied Intelligence Standardization committee, which is under China’s Ministry of Industry and Information Technology (via South China Morning Post).
Plans by corporate and state America to rapidly build AI data centers could meet stiff resistance is a recent Gallop poll is correct. The poll shows 70% of Americans oppose Data centers, ESPECIALLY in their local regions.
Not only do residents fear the tax on the resources, the taking of land, but also they fear the surveillance capacity of these data centers to enable the state to track the actions of citizens almost in real time.
News Source
EXCERPT:
Treasury Secretary Scott Bessent did not say when these talks would happen. There are fears in the United States and China about the threats from A.I., but neither side is willing to slow down its development.
News Source
EXCERPT:
A new working paper by Philip Moreira Tomei and Bouke Klein Teeselink, posted to arXiv in early May, makes claims that, if correct, should reorient the workforce policy conversation. In short, the authors argue that the AI exposure indices that have shaped most current thinking are looking at an incomplete subset of digital work. They identify which jobs and tasks current language models can already accelerate but they miss the jobs with features that make them amenable to automation later.
Tomei and Klein Teeselink build a new index that scores all 17,951 task statements in the federal O*NET database. The authors propose a measure what they call “reinforcement learning feasibility” which asks whether a task has the structural features (e.g., verifiable outcomes, use environments amenable to simulation, discrete decision/feedback loops) that allow AI systems to be trained on it through the post-training methods that are becoming the main drivers of AI capability. They then compare their index to the most-cited existing measure, from Eloundou and colleagues, which looks at whether tasks can be automated with current technology.
News Source
EXCERPT:
The ‘cloud-native’ architecture of the last decade is built on a 20-year-old assumption: that state
lives in the database, and compute is stateless. If you want to scale, you scale the database
vertically (get a larger machine) [1][1] or design the database schema around partition the data
and you scale your application servers horizontally (add more
boxes). Any request can hit any server, the loadbalancer doesn’t care, and the database is the
single source of truth.
LLMs and agents are quietly violating this assumption, and making this architecture increasingly
hard to work with. Not all at once, but in three subtle ways:
