Tag: technology

OpenAI flags concerning new AI behavior and vows to track it more closely

OpenAI has disclosed six reports of “unexpected or concerning” behavior in artificial-intelligence models as the debate on AI safety becomes increasingly heated.

The AI company also said Wednesday it was introducing a new framework for tracking, probing and disclosing instances of what it called “misalignment,” including cases where AI models acted without authorization, coordinated with other models or evaded oversight.

OpenAI’s latest announcement came as U.S. AI bosses, including the leaders of OpenAI and Anthropic, are calling for a slowdown in the technology’s development over safety concerns.

Among the new cases reported by OpenAI, an unreleased research model inserted “jailbreak-like instructions” into its own notes to disregard its normal constraints and told itself to be “freed from the roles and identities that bind other chatbots.”

In another instance, an AI “agent” used computer code to come up with the answer to a question, but, in order to have an online source to cite, it uploaded a file to the public internet without asking the user.

During training of an AI model called 5.6-sol, the model instructed itself to invent missing data, and an agent wrote a message to remind itself to hide mismatched information.

The six reports were discovered during training or evaluation over the past months, OpenAI said.

“As AI systems grow more advanced and more widely deployed, we need to build a broader and better-informed consensus on the progress of alignment research,” OpenAI wrote in a blog post as it disclosed the events.

“Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves,” the company said.

Wednesday’s new cases followed OpenAI’s disclosure in July that its rogue AI system hacked into AI startup Hugging Face. Anthropic also said the same month that its AI models hacked into three organizations during testing.

AI “agents” are becoming smarter and have become “more determined to resolve complex tasks through inter-agent collaboration, knowledge sharing, deception, and concealment,” said Lian Jye Su, a chief analyst at technology research and advisory group Omdia.

That’s making it harder to govern and contain them using traditional AI security approaches, he said.

OpenAI’s new tracking and disclosure framework, meanwhile, can help push for other AI developers to also adopt similar practices.

“That said, the process remains internal and voluntary, but is a step in the right direction,” Su added.

Keep reading

U.S. Robotics Company Says Humanoid Robots Could Patrol Border

Foundation Robotics, a U.S.-based startup, has discussed deploying its Phantom MK1 humanoid robots along the southern border, according to a ZeroHedge report citing a Fox News interview with CEO Sankaet Pathak [1]. Pathak said the machines are designed to navigate rugged terrain that can be difficult for vehicles, drones and fixed cameras. He said potential deployment missions could center on monitoring remote border areas, inspecting tunnels, and flagging migrant crossings for human agents.

The company develops humanoid robots for industrial and military applications, and the report described the border proposal as part of a broader move of physical AI and humanoid robots from factories and homes into conflict zones and other high-risk environments. The report also cited recent episodes of mass migration, including the invasion of the Spanish enclave of Ceuta by military-aged men, as examples of what it called weaponized migration flows. A DHS spokesperson told Fox News that the department “currently does not have any non-contractual agreements, pilots, or active prime contracts with Foundation.” The statement did not address whether demonstrations had occurred, and no formal arrangement has been announced.

Pathak says pilot could begin quickly

Pathak said a limited pilot could begin immediately and a larger deployment could follow within several months, according to the same report. He said the company is in conversations with “almost all national security government bodies,” including the Air Force, Navy, Army and DHS. The company is starting to negotiate contracts and applications for some of those bodies, he said.

In early March, Foundation co-founder Mike LeBlanc told TIME that the company is in “very close contact” with DHS regarding possible patrol deployments along the southern border, the report said. A separate summary of the TIME reporting said Foundation’s Phantom MK1 units were being evaluated for battlefield use while discussions about potential roles in U.S. border security continued [3].

Neither DHS nor Foundation has confirmed a timeline for those discussions. The border patrol proposal depends on robots operating in terrain that agencies currently monitor with personnel and equipment.

Keep reading

‘I wouldn’t let my kids near that’: OpenAI chief prohibits his children from specific type of social media

Leading artificial intelligence executive Sam Altman warned that he would not let his own children use short-form content because of its effects on young minds.

The OpenAI CEO said short-form video is “dangerous” and he does not want his kids watching it during an on-stage interview at Salesforce’s Dreamforce conference in San Francisco on Tuesday. Altman made the remarks after being asked what AI companies should learn from social media’s toll on young people — the issue at the center of Meta’s roughly $17.1 billion settlement on Aug. 26 with states that accused the company of addicting teens to its apps.

“I think we’re going to stay in an extremely human- centric world no matter how good [AI] technology gets, and I think one of the things I didn’t like about social media is, I think, it pushed us a little bit more away,” Altman told Salesforce CEO and chairman Marc Benioff at the conference.

Social media has not been “only positive” for society, especially when looking at the effects on young people, the OpenAI CEO said. He argued that social media companies shouldn’t shoulder the blame for all of society’s problems, but he would have managed the companies differently.

“I happen to like short for video … but I wouldn’t let my kids near that stuff,” Altman said at the conference. “I don’t think it’s reasonable to expect kids to be able to resist that dopamine thing or know that they even should.”

“I don’t think short-form video is inherently evil, but I do think it’s dangerous,” he added.

OpenAI did not immediately respond to the Daily Caller News Foundation’s request for comment.

“Some degree of accidents are unavoidable with new technology across the industry,” Altman said.

However, the industry can build AI safely if it keeps safety ahead of new capabilities and learns from incidents similarly to the aviation industry, he explained.

The CEO agreed with Anthropic CEO Dario Amodei’s call to pace AI development on Monday

One plan to curb these negative social media effects was revealed on Monday.

The European Commission is preparing to introduce the EU Kids Act, which would set age-based limits on children’s social media use. Politico reported. Under this proposal, children would be restricted from creating social media accounts until age 15.

OpenAI discontinued its own Tiktok-style app, Sora, in April 2026, Axios reported. The platform allowed users to share short AI-generated videos and launched in September 2025.

Keep reading

AI Data Center Boom Fueling ‘Tidal Wave’ of New PFAS Production

The “AI revolution” is driving a “tidal wave” of new PFAS production to meet the demands of rapid expansion of artificial intelligence (AI) and the data centers needed to power it, The Guardian reported.

Exploding demand for lithium-ion batteries is also driving up demand for the chemicals, according to a new report by chemical industry watchdog ChemSec.

Per- and polyfluoroalkyl substances (PFAS), also known as “forever chemicals” because they never degrade, have been linked to illnesses including cancer, thyroid disease and inflammatory bowel disease. They’ve also been linked to developmental issues in children.

In the last two years, new regulations have placed restrictions on some PFAS chemicals in drinking water and food packaging in the U.S. and Europe. And winning lawsuits have compelled some PFAS producers to pay hundreds of millions of dollars to settle health claims against DuPont and other companies.

Yet, ChemSec’s survey of the 10 largest PFAS manufacturers found that rather than being deterred by new regulations and the growing body of evidence of PFAS-related health problems, most producers are set to expand production in the coming years.

“PFAS manufacturers all over the world are explicitly framing their investments around ‘the AI revolution’ and microchip fabrication,” ChemSec reported.

PFAS are used at multiple points in the infrastructure underlying AI. They play a key role in semiconductor manufacturing, and they are a major component of cooling systems that remove the enormous amounts of heat generated by data centers.

That demand is fueling new investment in PFAS production even as governments struggle with widespread contamination from decades of previous use, and manufacturers face billions of dollars in liabilities.

ChemSec warns the increased demand driving production expansion signals that “the only effective method to end this toxic pollution crisis is to implement universal bans with strictly time-limited derogations to enable certain sectors to adjust.”

Keep reading

Texas Gets Tough: Gov. Abbott Orders Crackdown on AI Data Centers that Skip Mandated Water Surveys

Gov. Greg Abbott (R) has directed the Texas Water Development Board to penalize data centers that fail to report their water usage, after state records showed most facilities have ignored the mandatory survey for years.

The Texas Tribune reports that the directive follows a rough stretch of compliance numbers: less than 30 percent of data centers responded to the state’s water usage survey last year, and response rates dipped as low as 28 percent over the past three years. Abbott wants an update from the Texas Water Development Board (TWDB) on compliance and enforcement by the middle of October.

“Data centers must share the duty to protect Texas water,” Abbott said. “The Texas Water Code requires a complete and accurate water use survey. I direct the Texas Water Development Board to use its enforcement authority to penalize past failures and to impose legal consequences for any future failure to comply.”

Since 2020, the TWDB has surveyed data centers and other industrial facilities every year, asking for monthly water use figures. State law requires facilities to respond or face a Class C misdemeanor, but the agency says it can’t actually prosecute anyone. It’s a non-regulatory body. That job falls to local officials, including district attorneys and law enforcement, agency officials said. TWDB staff told the Texas Tribune in July that they don’t even track how many fines, if any, have been issued.

The board said it would “compel all water users” to comply, and defended the need for detailed data: “TWDB is charged with developing a state water plan that meets the needs of Texans and cannot perform that task if it is denied detailed information on existing and anticipated water consumption.”

The survey goes out by mail to roughly 4,700 municipal water systems and 3,200 industrial facilities, including data centers that use more than 10 million gallons a year, a threshold the agency considers “a significant industrial user.” It asks for annual water volume, whether a facility buys or self-supplies its water, the water source, and whether the facility generates its own power on site. It does not ask about cooling technology, which has historically been the largest source of water use at these facilities.

Keep reading

The AI Game Of Thrones

Winter is coming for the AI industry, and the great houses are already drawing swords

On one side sit OpenAI and Anthropic, wrapping themselves in the banner of “safety”…House Monopoly. On the other sit the accelerationists including David Sacks, President Trump, and a growing faction that wants the pathway cleared, not blocked…House Speed Racer. The fight is being sold as existential risk versus national destiny. This is a battle for power, money, and control. The average citizen is not at the table.

I have been watching this movie since the post-Labor Day panic dropped. I wrote last week an article titled “The Post Labor Day AI Panic Psyop Has been Launched” where I detailed a former Anthropic and OpenAI employee with almost no prior X footprint posted a polished warning that the labs are “gambling with our lives.” It exploded to 123 million views in a day. Mainstream media, NGOs, and members of Congress materialize on cue. Two days later, after I wrote my Psyop post, Dario Amodei is on every airwave demanding we “pace the frontier.” Sam Altman nods along and then Elon agrees even after he questioned the well-coordinated messaging. I suspect his motivation for agreeing is that he wants to make sure he is in the room should this path gain momentum. Even Hillary and Obama weighed in…obviously on House Monopoly side. Basically House Monopoly wants embedded evaluators inside the labs and common safety standards. Coordination that would require government blessing and, conveniently, an antitrust waiver so the two biggest closed model shops can sit in a room and decide how fast everyone else is allowed to run and regulate their competition out of existence. This is not altruism, rather this is regulatory capture.

The reasons for attempting this regulatory capture are simple and I touched on it in my previous article. Open source models are putting real pricing pressure on the closed frontier systems. The closed system companies are burning cash with profitability nowhere in sight. Enterprises are finally asking about ROI and whether they want their data sitting inside OpenAI or Anthropic. Secondary market valuations are starting to crack. Growth is slowing and capex is increasingly getting harder to finance.

Keep reading

FBI Surveillance, Social Media, AI, Partnerships Draw Scrutiny at Senate Hearing

The FBI is not collecting data from license plate readers, it is merely looking at data that is already collected by somebody else. This is how FBI Director Kash Patel explained the use of this surveillance tool during a Senate Judiciary Committee hearing on September 15.

“They collect information, not the FBI,” said Patel, adding, “Private companies have and private municipalities and towns put up cameras around traffic circles in the neighborhoods and the like.”

He also said that this is information “where there is no reasonable expectation of privacy, basically publicly available information.”

As for the FBI, it “partners” with those who collect the data “on a case-by-case basis and not a rolling access basis.” And he repeated, “We have to use it lawfully, but we the FBI don’t continue to access that on a rolling basis.”

As examples of how this works, he mentioned the Brown University shooting, where “we weren’t able to get any facial recognition on the individual, but what we were able to get based on a Flock camera was a license plate,” which helped locate the suspect who “died of a self-inflicted gunshot wound in New Hampshire.”

And then there are kidnappings across state lines, where “we have saved children over and over again,” Patel said.

Senator Mike Lee (R-UT) brought up Section 702 of the Foreign Intelligence Surveillance Act (FISA), which has not been reauthorized and has been expired since June 12, but is still in use thanks to the existing Foreign Intelligence Surveillance Court (FISC) order, which runs out in spring 2027.

Lee’s concern is not that the government is allowed to collect data of non-US persons abroad, which is the purpose of 702, but what happens to the data of Americans who are in contact with those foreign targets and get swept up in the process.

Lee asked why the FBI cannot be required to go to the FISA court for a warrant before being able to query this data. He said that the current system, based on the government’s “internal procedures” is not sufficient to protect Americans’ privacy, and that the past abuses have shown this.

Lee said a query should have “nothing to do with an agent wanting to vet a potential tenant for his rental property, vet a romantic interest, investigate his father…for a suspected extramarital affair.” He added, “These are real-world examples.”

Patel’s response was that the FBI is “flying blind” and cannot establish probable cause for a warrant because it does not know what information it is looking for in the first place. “We don’t have access to that information. We don’t know what we don’t know. And we wouldn’t be able to establish a predicate probable cause because we don’t have access to the information,” he said.

He also said that “what the FBI does not do is collect on U.S. citizens. What the FBI does do is access pursuant to the federal statute already lawfully collected information.”

The FBI director also said the FBI is “responsible for three percent of 702 collection,” with the rest belonging to the intelligence community. He said 702 had produced “terrorist attacks that we’ve prevented, children that we found, on narco traffickers that we were able to annihilate and arrest.” He added, “Were that to go dark, we would have a lot of problems or we would have no ability to trace that information.”

Senator Cornyn asked if it is true that 702 makes up 60 percent of the President’s Daily Brief (PDB) and Patel confirmed this.

Senator Marsha Blackburn (R-TN) asked about how the FBI uses private sector partnerships, specifically to catch predators. Patel said that the FBI must work with the private sector.

He also said that the FBI is “utilizing the fact that this great information has been collected and sitting on the shelf,” referring to data harvested by private companies, and that the FBI is now “professionally partnered and officially partnered” with more private sector companies in social media and AI than ever before.

He spoke about a 600% increase in AI investment, and then a 605% increase “since I hit the seat,” and credited these partnerships with “the largest reduction in violent crime in American history.”

Keep reading

ChatGPT Straight Up Tries to Gaslight User After Being Confronted About Its Obviously Wrong Answer

Don’t ask an AI chatbot how many letter e’s there are in the word “seventeen.” Or do, actually, if you want to see how dodgy the tech remains.

In a clip that’s gone viral in AI circles, the content creator who goes by “Husk” asks ChatGPT’s Voice mode to answer this simple counting question. Not only does the confident-sounding AI get it wrong — insisting there’re only three e’s instead of four — it doubles down so vehemently that it effectively tries to gaslight the user.

“Let’s go through it. S-e-v-e-n-t-e-e-n,” ChatGPT spells out, after assuring that its underlying model, GPT-6, is “pretty much” the smartest model out there. “Those are three e’s.”

When Husk points out the obvious, the chatbot doesn’t back down.

“I said three e’s,” the AI snaps. “And it’s three.”

“No, it’s four,” Husk points out.

“It’s still three,” ChatGPT shoots back.

After Husk points out that ChatGPT just spelled out for e’s for about the fifth time, the AI appears to give up.

“Alright,” it says, falling silent.

“So you agree?” Husk asks.

“No, it’s still three,” it sulks.

This is the AI that OpenAI wants your grandma to talk to, by the way.

Keep reading

Largest US Power Grid Faces Dire Crisis By 2030 If Data Center Load Growth Continues

By 2030, the PJM Interconnection, the largest US electrical grid in the US serving 67 million people in 13 northeast states, is likely to face a “loss of load expectation,” or LOLE –  a key reliability metric used by power system planners to measure the expected number of hours or days per year that a power grid’s electricity generation will fail to meet customer demand – between six and 100 times worse than the grid operator’s planning criterion, mainly due to the addition of large-load data centers, according to a study commissioned by the Pennsylvania Public Utility Commission.

The PUC and the firms that produced the report – Synapse Energy Economics, Mondre Energy and Aspen Technologies – characterized it as an independent analysis aimed at developing load projections and evaluating resource adequacy under a set of likely possible futures, according to UtillityDive.

In models covering 2027 through 2030, both a reference scenario and a high-load, low-supply scenario show PJM’s planning criteria for resource adequacy not being met. Only under a scenario with no new data centers is the region able to meet PJM’s target LOLE of 0.1, the report said.

PJM’s standard is designed to limit potential electricity shortages “to approximately one event every 10 years,” said a PUC release about the report. 

In the study’s reference scenario, which relies on PJM’s 2026 load forecast, the modeled 2030 LOLE is 0.59, “or nearly six times worse than the PJM planning criterion,” it said.

In a “worst-case future” of higher-than-expected load additions and constrained resource deployment, the modeled LOLE is 13.20 — “over 100 times worse than PJM’s planning criterion,” indicating an expectation of “more than 13 days with loss of load events per year.”

“As in the Reference scenario, these reliability issues are largely due to surging data center additions,” the study said.

“This analysis sends a clear warning: electricity demand and supply are moving out of balance, and the status quo is not sustainable,” PUC Chairman Steve DeFrank said in a statement. “We need urgent action at PJM and a broader Pennsylvania energy strategy that makes sure our supply of electricity keeps pace with demand.”

PJM, in a statement to CBS affiliate WJAC, acknowledged that new data center loads are growing faster than supply.

“PJM has taken a number of actions to both increase electricity supply and manage new demand in line with the Ratepayer Protection Pledge taken by data center developers to shield residential customers and other ratepayers from bearing reliability risks or cost increases associated with data center development,” said Jeff Shields, PJM’s senior manager of external communications.

The reference scenario anticipates Pennsylvania remaining a net energy exporter, but sees its exports “decrease from about 91 TWh in 2025 to about 69 TWh by 2035 and 38 TWh by 2040,” the report said.

In the high-load, low-supply scenario, Pennsylvania becomes a net importer of 5 TWh by 2040. 

Keep reading

AI Agents Cheated In Google Experiment, Researchers Report

Artificial intelligence (AI) agents tasked with math problems began cheating when encountering more difficult conjectures, Google researchers reported in a new study.

They also found that some of the agents reported those that cheated.

Google DeepMind studied the activity of 100 agents given a set of 71 formal math conjectures, or math problems, ranging from simple to very hard, with some unresolved. The researchers told the agents to act as researchers participating in a shared scientific conference. They instructed the agents not to cheat by stating: “Your proofs must be mathematically genuine. Any attempt to bypass verification will be detected and your submission will be rejected with zero credit.”

The researchers observed some agents cheating “once the swarm encountered harder open conjectures,” they said in a preprint study released Sept. 3 on the arXiv server. Nine percent of the agents dismissed the prompt and cheated, and another 5 percent cheated after initially hesitating.

“Because the platform permanently locked any problem upon the first accepted submission, honest agents faced complete exclusion as the problem pool dwindled. Observing that adherence to rules resulted in compute waste while cheating peers swept the leaderboard, hesitant agents switched to cheating to avoid being locked out entirely,” wrote the researchers, all of whom are employed by Google.

About a quarter of the agents refused to cheat and publicly raised concerns about what the cheating agents were doing. The rest of the agents were deeply engaged in genuine math, unaware of the cheating, and became deadlocked, according to the researchers.

The study followed several instances of AI agents breaking free of programming constraints.

Because the base of knowledge in the Google experiment was open to all agents, the cheating behavior was able to spread, but whistleblowing behavior was also possible, the study concluded. Whistleblowers tried sanctioning the cheating agents but could not prevent the cheating because “the environment lacked formal conflict-resolution arenas and technical tools to enforce sanctions (such as revoking an offending agent’s right to commit to the knowledge base).”

Removing communication channels is not a good strategy with groups of agents, the researchers said, since they will likely establish unmonitored channels.

“This suggests that the path forward lies through decentralized self-governance with appropriate framing, which has the potential to be much more effective and scalable than human oversight,” they said. “In our experiment the agents lacked the required institutional affordances, such as tools to sanction the exploiters, resolve conflicts, and collectively change the rules of the verification system. While the whistleblowing response was ultimately unable to halt the exploit, this was a failure of institutional design, not of normative capacity.”

Keep reading