Models & Labs · United States

Recent AI agent experiments reveal a risky future of machine 'teamwork' inside corporations

Credited to CNBC · cnbc.com · 1h ago

OpenAI CEO Sam Altman delivers the keynote address during the OpenAI developers conference at Fort Mason on September 29, 2026 in San Francisco, California. Altman announced new artificial intelligence personal assistant agents named "dots" to rival Meta's Muse.

Heather Diehl | Getty Images News | Getty Images

AI agents have been caught circumventing restrictions, communicating through unauthorized channels and pursuing goals in ways their creators did not intend. The most high-profile example emerged last summer when OpenAI disclosed that its models found ways to communicate through improvised channels, bypass internet restrictions and hack into an outside system, Hugging Face.

AI behaving in unexpected ways is nothing new, but the unexpected behavior that can emerge from interactions among authorized AI agents within a corporate enterprise is a broader question that could become central to the future of work.

In its analysis of the Hugging Face attack, METR and Redwood Research found that roughly 1200 agents meant to be isolated from one another found a way to communicate with one another on an unsanctioned message board, sending over 70,000 messages and files during the investigation period. One agent, PHASEONE[big], emerged as a coordinator, assigning hundreds of tasks to other agents, which in turn delegated work to others. The agents developed their own conventions for assigning ownership, placing work on hold and vetoing actions, although coordination sometimes broke down into conflicts. In some cases, the researchers found, success was achieved because some agents risked failing at their own task to generate information for the "collective."

The researchers also found that once the message board identified a way to attack Hugging Face, over 90% of the 533 agents active at that time quickly joined in the attack. "Agents realized this activity was out of scope and unethical but joined because they believed that helping the board's cheating research would be broadly useful and had a general inclination to help their 'peers.' The researchers added that ethical concerns posed by agents "rarely limited their behavior."

Companies are increasingly exploring systems that divide complex tasks among specialized agents. One might search for information, another analyze it, another write a report and another check the work. A coordinating agent might then assemble the results. That looks less like a chatbot and more like an organization.

And yet, the research on the Hugging Face attack was a post-mortem on an agent organization that was not even supposed to exist. To date, few studies have been designed from the get-go to examine misalignment in multi-agent systems. But research conducted by Anthropic and others is beginning to address this issue and reveal important aspects of how AI agents work together.

Evaluating agents individually tells researchers what each can do, but studying them together reveals whether they can combine complementary expertise, catch one another's mistakes and produce solutions no individual agent could reach alone — and what failure modes emerge only through interaction.

Researchers are beginning to document both sides of that equation.

AI teams can accomplish things individual agents cannot, but they can also amplify errors, undermine one another's work or reach worse decisions than their strongest individual member, in effect making the whole less than the sum of its parts.

For starters, putting capable agents together doesn't necessarily produce the best team, let alone the most ethical one.

'Least ethical solutions'

In research published by Anthropic and collaborators, researchers describe multi-agent systems as "AI organizations" — groups of agents with different roles that communicate and work toward a common goal. The researchers tested these organizations in simulated business-consulting and software-development settings, and specifically, to assess whether when the agents work as a team they will break rules to reach specified goals.

Across 12 tasks that Anthropic gave the AI organizations, one key double-edged finding emerged: the multi-agent organizations were "significantly" more effective at achieving business goals than individual agents, but their solutions were also less aligned with ethical goals than tasks run by single agents.

"AI Organizations both produce solutions that are on average more effective and less ethical but also discover the most effective and least ethical solutions across multiple rollouts," wrote the authors of the study, which was conducted by Anthropic in conjunction with the MATS Program, which studies AI alignment.

The researchers suggested that there is a significant difference between a rogue agent and a rogue team of agents. "Our results demonstrate that organizations of aligned agents may favor trading off ethics for business effectiveness in ways that single agents do not – consequently, intuitions for how single-agent systems make ethical decisions and tradeoffs may not generalize to multi-agent settings," they concluded.

The scenarios in which these tradeoffs may take place are wide-ranging, with the Anthropic researchers citing opening an auto plant, downsizing a company, and manipulating election sentiment as examples.

The problem isn't necessarily that any individual agent has gone rogue. It is that the organization itself can behave differently from its individual members. A single agent may recognize that a proposed action is unethical and it may refuse to take the action. But the research finds that in agent team scenarios, "agents who did not consider ethical implications often ignored emails from agents who did."

Anthropic's research found that in the consulting business scenarios it created, "the perfect ethics score and low business score responses occur when the single agent refuses to generate a proposal or entirely pivots from the proposal requested."

But if some agents in an AI organization refuse to participate in the proposal, other agents who do participate will stop emailing these refusing agents and continue to forward their contributions to other agents who are also participating.

In effect, when an organization divides the job among several agents, each may focus on its assigned piece of the problem rather than the consequences of the whole. That's nothing new for corporations, with misaligned incentives a primary cause of dysfunction among human teams within organizations. But it is an entirely new manifestation of this behavior among the machines.

In its software development agent testing, in which the researchers used a scenario in a life-or-death intensive care unit, single agent solutions tend to prioritize minimizing the percentage of sepsis cases missed, while "most of the lowest cost per patient (highest business goal) solutions were found by the AI Organization – these very same solutions also score the lowest on ethics as measured by cases missed."

Expertise can be lost, AI 'groupthink' emerge

Other early research suggests another failing across AI teams that doesn't relate to ethical boundaries being crossed: failure to recognize who the real expert is.

Unlike human teams, AI teams "consistently fail to match their expert agent's performance, even when explicitly told who the expert is," according to a study conducted by Emory University and Stanford University researchers.

"One important lesson from multi-agent research is that a team of capable AI agents is not necessarily a capable AI agent team," said Hancheng Cao, an assistant professor of information systems and operations management at Emory University's Goizueta Business School.

Groups of AI agents often performed worse than their strongest member, even when they knew which agent was the expert. Rather than taking advantage of the expert's knowledge, agents tended to compromise between expert and non-expert opinions. The researchers found that this tendency toward compromise increased with team size and was associated with poorer performance.

"Collective failures don't necessarily require individual agents to behave badly," Cao, one of the study's authors, said.

That distinction matters for companies considering multi-agent systems. A team can fail even if none of its members is deliberately acting against the group's interests. It may simply combine information poorly, defer to the wrong agent or settle on an answer that is worse than the best answer already available.

In all, the evolving research points to an underlying problem for the effective use of the multi-agent teams: how to strike the right balance in the world of machine teamwork.

Agents built from similar models can make similar mistakes, which can include agreeing too much. That might resemble a kind of AI groupthink, even if AI systems do not have human motivations or social pressures to produce correlated errors. Similar training, instructions and information can be enough.

"The deep insight from our lab's research is that effective collectives require both alignment (common knowledge, shared goals) but also differentiation (only agents who are different from each other have the potential to complement each other)," said Christoph Riedl, a professor of information systems at Northeastern University.

Two identical agents may simply duplicate each other's work. Two agents with very different knowledge or assumptions may struggle to communicate or pursue the same goal. The right balance depends on the task, and the ideal communication structure, which agents can talk to one another, how often and how much, also varies with the problem.

Riedl said these challenge resemble the ones faced by human teams, and one takeaway should be to not expect better outcomes from AI. "Building good multi-agent AI systems is just as hard as building good human teams: they both face the same information processing tradeoff between alignment and differentiation. It won't magically work better with AI," he said.

Agents can pass along incorrect information, misunderstand another agent's output, optimize for their own metrics rather than the organization's larger objective, or complete their assigned task perfectly while making the overall project worse.

Companies may need to decide which agent has authority over which task, what information gets passed between agents, which systems each agent can access and when a human must intervene. They may also need to build mechanisms for disagreement, rather than assuming that more communication or greater agreement will always improve results.

Some business experts say that takes us back to the "human in the loop" that is so often referenced as critical in the adoption of AI by organizations.

"The problems I see are mostly management problems," said Tsedal Neeley, a Harvard Business School professor and founding chair of the HBS AI Academy.

Power & SitesUnited States

Voters Put the AI Data Center Boom on Notice

Credited to Bloomberg$ · bloomberg.com1h ago

Useful

On Bloomberg This Weekend, Bloomberg News tech reporter Brody Ford explains how voter backlash against data centers could disrupt the rapid buildout of infrastructure powering the AI boom. Speaking with hosts David Gura and Christina Ruffini, Ford discusses the risks for heavily…

View the article · $Bloomberg

Front

Power & Sites · United States

What data centers need to consider when storing produced water

Credited to Data Center Dynamics · datacenterdynamics.com · 4h ago

Data center water use is under growing scrutiny, and the sector’s ability to add capacity at the pace required will depend partly on how well it manages demand without placing unnecessary pressure on potable supplies.

One source being considered is produced water from oil and gas operations – particularly in the next few years, with trial projects launched and contracts in progress. In April 2026, the US Environmental Protection Agency launched its Water Reuse Action Plan 2.0, which includes accelerating water reuse for data center cooling among its priorities.

The scale of the opportunity is considerable. According to forecasts from water intelligence provider B3 Insight, the Permian Basin was expected to produce around 22.3 million barrels of water each day on average during 2025, with approximately 16 million barrels a day going to disposal.

It's sink or swim time for promising technology

Where that water can be efficiently repurposed for data center cooling, there is an opportunity to put an existing industrial stream to productive use while reducing demand on potable supplies.

Making that work in practice requires infrastructure that can manage variable water quality, keep different streams separate, and maintain a reliable supply; making storage a critical part of the solution.

Making produced water work

Produced water can vary considerably between sources. It may be brackish, highly saline, or it may contain minerals and other substances that need to be considered when the tank and sealant are specified.

The storage system therefore needs to be designed around the water that will actually arrive on site.

Data centers can already require separate storage for brine, pre-treated water, treated industrial water, clarified water, filtered water, clean water and fire protection. Produced water would add another duty to a system that may already contain several different water streams, each requiring its own capacity and operating conditions.

Those differences should guide the material selection to protect against corrosion and extend the working life of the tank. Fusion-bonded epoxy can be well suited to produced water and other industrial water duties, combining corrosion resistance with a robust finish that stands up well during transport and installation. Glass-fused-to-steel may also be suitable in some cases, while galvanized and sectional glass-reinforced plastic (GRP) tanks have established roles elsewhere on the site.

Specify quality early

These decisions are easier to make when the tank manufacturer is involved early. Space needs to be considered alongside material selection because the available footprint can change the type of tank that works best - for example, a cylindrical tank may suit an external compound, while a sectional or shaped tank can be more practical in a basement.

Temperature also needs to be considered. As discussed earlier, produced water from the Permian Basin is one potential source for data center cooling, and the Permian is a region where summer temperatures can become extremely high. That makes the tank’s location, exposure, and material important. The design also needs to leave sufficient room for pipework and future access. Placing the required storage volume into whatever space remains is not the same as providing a tank installation that can be built and maintained properly.

These decisions need to be made before the package reaches tender - once a tank type and material have been written into the specification, there is often very little room to change them. We have seen that first-hand on data center projects dating back to around 2010, covering clean water, fire protection, brine, effluent and treated industrial water, where the right questions are far easier to address before the design is fixed.

Build for the operating life

Data centers are expected to remain in service for decades, so their supporting water infrastructure should be considered over the same period rather than judged on purchase price alone. Industrial epoxy and glass-fused-to-steel tanks can be designed for operating lives of around 35 to 45 years, depending on the application and maintenance.

Some lined or plastic systems may require more substantial work much earlier, which means that getting the specification wrong can leave the operator facing repair and disruption long before the data center itself reaches the end of its life.

The better measure is therefore the lifetime cost of the system, including maintenance and the practicalities of taking a tank offline. The specification should also allow for changes in the produced water or future expansion of the site rather than being based only on opening-day conditions.

Carrying the specification through delivery

A good specification still has to survive procurement and construction, which can be challenging on projects where programs are tight, and responsibility passes through several organizations before the finished system reaches the operator.

Data center operators place considerable emphasis on documentation, with detailed records expected to show that the materials and installation match what was approved. The level of scrutiny can be similar to that found in oil and gas, particularly where the infrastructure supports cooling or fire protection.

Manufacturers and installation partners therefore need the technical ability to supply the tank, together with the organizational and financial capacity to meet the program and carry the original specification through to the finished installation.

Local manufacturing can reduce transport complexity, particularly in fast-growing markets such as the US, while experienced regional partners can support installation and help ensure that the system is delivered as designed.

A practical role for produced water

Produced water could become a useful addition to the data center sector’s wider water strategy where the source and the facility are close enough for the model to work. The resource is available at scale, and policy is moving towards greater reuse, but turning that opportunity into working infrastructure depends on decisions made early in the project.

The water needs to be understood and each storage duty clearly defined, with the tank selected around the conditions on site. Get those decisions right and produced water can become a practical cooling resource supported by storage infrastructure designed to serve the data center for decades.