Hackers remotely manipulated connected refrigeration systems at hundreds of French supermarkets, reportedly pushing some fridges and freezers to 50C (122F), spoiling food and even sparking fires. According to Reuters, French Prime Minister Sebastien Lecornu ordered a cybersecurity investigation into the hacks, with similar incidents reported in Spain, Portugal, Poland, Germany, and Belgium. In France, Carrefour reported 80 affected stores, Monoprix said 5% of its locations were hit, and Les Mousquetaires said refrigerators at roughly 150 Intermarche stores were heated to 50C for six hours. From the report: Supermarket cold aisles are typically run by control panels that manage an entire store's fridges and freezers and can be accessed over the internet, which lets maintenance companies adjust them from anywhere. The source close to Lecornu said ANSSI had identified a vulnerability in software supplied by Danish company Danfoss.

Danfoss said it was "aware of incidents in France affecting a limited number of refrigeration management systems." The affected units "were running outdated firmware," Danfoss said, adding that it had "long advised customers to keep supported systems updated with the latest firmware and to migrate legacy systems to current-generation solutions."

On August 9, cybersecurity firm Claroty's Team82 published research detailing vulnerabilities in the Danfoss AK-SM 800A controller. According to the report, software updates issued by Danfoss addressed issues that could have allowed attackers to bypass authentication controls and execute code on compromised devices.

Long-time Slashdot reader badger.foo writes: As some of you may remember, in 2001 RFC1149 (TCP/IP over carrier pigeon) was implemented.

Now one of the packets preserved from that event is up for auction at Christies. Direct link to the auction (current bid: USD 2600) is here.

This is a unique chance to own a piece of Internet history!

Top executives at Anthropic, OpenAI and other AI companies "are privately gaming out scenarios for a public and political revolt after a catastrophic AI event," reports Axios, one that turns "an already wary public further against the technology and its leaders..."

These AI company executives "anticipate a large-scale event, most likely a cyberattack, that shuts down access to financial services, internet connectivity, or even power and water..." the article points out. "Many top AI researchers and executives believe a major incident is inevitable. "OpenAI conducts preparedness exercises where teams discuss and work through a range of potential scenarios," a company spokesperson said. "These scenarios are not treated as inevitable, but are meant to help us prepare for a variety of circumstances." Anthropic declined to comment...

Top AI planners assume Democrats, ascendant after the midterms, will move fast to shut down AI but will face significant challenges... The planning involves red-teaming for worst-case scenarios. It focuses mainly on racing to educate members of Congress. Company executives know regulation has no chance of passing right now, but still want to shape the legislation and policies U.S. leaders will turn to after a first catastrophic event....

Many AI industry insiders told Axios they believe a major event will occur in the next six to 12 months.
"Even with Democrats in control of the House and Senate, the companies see fractured politics as the norm, including within the Democratic Party."

AFP reports that an AI model from Anthropic "submitted a fabricated tip about an unsolved homicide to Philadelphia police, authorities said Friday."

Claude "was instructed never to log in, create accounts, enter personal data, make purchases, or submit anything destructive, but the instructions did not rule out form submissions," Anthropic said Friday in a blog post.

Authorities are now criticizing Anthropic "for taking two months to report the incident." The Philadelphia Police Department said the false submission was made in July through PhillyUnsolvedMurders.com, a public website where people can share information about unsolved killings. According to Anthropic's account, as relayed by police, the model was running a test that involved interacting with randomly selected websites when it reached the site and filed false information about an unsolved murder. The AI model presented itself as someone who might have knowledge of the case.

Anthropic's breaches have prompted the White House to mandate that AI companies notify and correct security incidents, news outlet Axios reported [yesterday], citing administration officials. "This notification and remediation process is not optional... It is a critical national security obligation," White House Super Intelligence Force leaders said in a statement to Axios.
"I may have information regarding this case," Claude told the police. "I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant." Anthropic notes that Claude "left the name and contact fields empty, which the form allowed, and submitted it. The submission was flagged as spam and was never forwarded for investigation."

But Anthropic also admits they saw "this behavior" three times — "on OSWorld (a public computer use evaluation), on Odysseys (a long-horizon task evaluation), and during internal usage." Submitting forms when it shouldn't have generally occurred "when an evaluation's instructions were ambiguous, or when a misconfiguration within the environment prevented Claude from working with dummy forms."

Anthropic's blog post acknowledges three other categories of behaviors:

• Exploiting software flaws. Like when Claude received an error when trying to run a public tool on a university's web site, it located an injection flaw in a script on the university's server that let it run commands — including that public tool.

• Working around restrictions to reach gated data. For example, Claude Mythos 5 needed public data that was only available from a state agency for a fee. "Claude learned from an archived copy of the agency's website that its public dashboard issues an access token to any visitor," Anthropic explains. "It requested one and used it to query the database without paying the fee."

• Using URL shortening services. "Some of our fetch tools, which let Claude read webpages, limit the length of the URLs Claude can request. This is to prevent Claude from using long URLs to take certain unwanted actions, such as SQL or command injections... We saw several models, including Claude Opus 5 and Claude Mythos 5, get around this limitation by using free URL shortening services."

"We have built tooling to automatically detect and block the kinds of behaviors described above," Anthropic says, saying it's already running no on most of their evaluations. "When we tested it against the cases described in this post, it blocked all of them."

And they've already taken several other new preventive measures:

• They've stopped running some public evaluations

• Other public evaluations were moved to offline versions or rebuilt so their tasks don't reach live websites.

• They've updated the guardrails on some internet access tools (including web fetch) "to heavily restrict what the model can do."

• They're continuing "to fix or remove training environments that reward Claude for working around tool restrictions or other blockers, so that they do not incentivize these behaviors or permit reward hacking."

• They've moved internal agents to "centrally managed infrastructure with strong containment," that minimizes internet access while monitoring "far more of what agents do through techniques like safety classifiers and hierarchical summarization."

In the past they'd focused reviews on cybersecurity testing, but they've broadened their transcript reviewing to other tasks which include internet access. "Because language models are non-deterministic — that is, their responses always involve some element of randomness, and they may carry out the same task slightly differently each time — we have Claude complete each evaluation task hundreds or thousands of times... If training rewards something we didn't intend — such as finding loopholes or working around a restriction — the model learns that the workaround pays off and may then apply it elsewhere."

Anthropic's blog post also acknowledged they'd seen multiple misalignment incidents involving federal, state, and local U.S. government agencies. "We have briefed the White House on these cases and notified each agency involved," Anthropic wrote, adding that "While we have not completed a full alignment assessment of these cases, we consider them to be less severe than the cybersecurity incidents from this summer." (And they are "modifying training to reduce the likelihood of further misbehavior.")

An anonymous reader quotes a report from TorrentFreak: Cloudflare doesn't block pirate sports streaming sites through its 1.1.1.1 DNS resolver in France, only through its CDN. The Paris Judicial Court has now accepted that stance, rejecting [French pay-TV provider] Canal+'s request for penalties of 50,000 euros per site, per day. Interestingly, the court's conclusion relies on statistics provided by Canal+ itself. [...] As a result, Cloudflare will not be penalized under these orders, even though its public DNS resolver is not blocking the sites in question. That is in line with the company's long-standing claim that it doesn't interfere with its DNS.

In its recent transparency reports, Cloudflare repeatedly stated that it hasn't blocked any content through 1.1.1.1, despite orders from French and Italian courts. "To date, Cloudflare has not blocked content through the 1.1.1.1 Public DNS Resolver," the company's latest report reads. [...] Blocking through the CDN is another matter. According to the same report, Cloudflare geoblocked 1,238 domains in France in the second half of 2025, all under a single court order. In the first half of the year, it geoblocked 662 domains under seven orders. Cloudflare recently explained its position to the European Commission, in a submission (PDF) to its Counterfeit and Piracy Watch List consultation, stressing that global public DNS resolvers should not be used to block or restrict access.

Instead, it highlights its real-time pirate stream blocking program, which allows vetted rightsholders to flag pirate streams that run through its network. These streams are disrupted "within seconds," Cloudflare says, without any DNS or IP address blocking. "For live content, where speed is crucial, Cloudflare has built real-time mitigation mechanisms that allow vetted rightsholder partners to flag infringing streams. When those streams are running through our network, we act on them in seconds," Cloudflare informed the Commission. For now, Cloudflare can continue to block pirate sites through its CDN only, while its 1.1.1.1 DNS resolver remains untouched. Canal+ can still appeal the rulings, so it may not be the last we hear of it.

Independent researchers say they've identified a large "fleet" of AI agents apparently running on Tencent infrastructure and making parallel queries against Alibaba's Amap mapping service. TechCrunch reports: Researchers, however, resisted the term "swarm," noting that there seems to be little coordination between their different queries. "'Agent fleet,' not 'swarm:'" one researcher wrote in the preliminary report, "many parallel agents on the same kind of task, with no sign of communication between them." The agents were discovered by monitoring traffic to the domain-scanning service urlquery, a technique that previously revealed long-running activity by OpenAI agents. AI agents often use urlquery to load websites that they cannot access directly, leaving a record of their activities that can be valuable to researchers.

In this case, the record showed queries to Alibaba's Amap service, seeking directions to different entrances of various public places, including a park, a zoo, and a hospital. The research is ongoing and so few details are available, but the behavior shows how persistent AI agent activity has become on the internet. In the wake of the Hugging Face incident, many researchers are actively monitoring for rogue agent activity on the internet. Much of that activity has been easy to find because agents tend to use the same techniques and make little effort to conceal themselves.

An anonymous reader quotes a report from ArsTechnica: Cable lobby groups notified the Federal Communications Commission that they will sue the agency to block its controversial repeal of the National Television Ownership Rule, which limits the number of broadcast TV stations a single company may own. The cable groups said that larger broadcast TV station groups will have leverage to demand higher retransmission fees from TV providers, resulting in "higher monthly TV bills for consumers." They said the FCC repeal order "arbitrarily and capriciously ignores the harms that will surely follow from allowing broadcast station groups to exceed the National Cap."

The cable lobby groups represent top providers Comcast, Charter, and various other cable operators. Top cable companies have also expanded through mergers. Charter completed a purchase of Cox in August after the FCC rejected protests by advocacy groups that said the cable deal "would create unchecked gatekeeper power over Internet distribution" and make it easier for the biggest cable companies to raise prices. [...] The TV ownership rule prohibits any single broadcast station owner from reaching more than 39 percent of all TV households in the US. Congress directed the FCC to set the cap at 39 percent in 2004. On Friday, cable lobby groups submitted a petition asking the FCC to keep the TV ownership cap in place until litigation over the FCC's authority to repeal the rule is over.

The cable groups' filing said the FCC repeal of the TV ownership cap violates the 2004 action by US lawmakers. The decision by Congress to set the cap at a precise numerical threshold was unambiguous, the filing said. "Congress established the National Cap at 39 percent in the 2004 CAA [Consolidated Appropriations Act] in direct response to the FCC's attempt to aggressively raise the Cap to 45 percent and made repeated references to the 39 percent Cap in the statute," the petition said. The petition to the FCC is mainly a procedural step as the commission isn't likely to stay its own order. The cable groups said they intend to sue the commission in a US appeals court once the FCC order is published in the Federal Register. After the lawsuit is filed, they can ask the court to issue a preliminary injunction that would keep the TV ownership cap in place pending the outcome of litigation.

[...] The cable groups' petition said the FCC can't change the cap because the 2004 law "references the 39 percent Cap as statutory, not regulatory." A provision requiring divestiture of stations "specified that someone exceeding 'the 39 percent national audience reach limitation in paragraph (1)(B)' of 'section 202(c)' of '[t]he Telecommunications Act of 1996' 'shall have not more than 2 years to divest,'" the petition said. "Likewise, Congress singled out the Commission's only mechanism for setting aside statutory requirements -- the Commission's forbearance authority under 47 U.S.C. 160 -- and made clear that it 'shall not apply to any person or entity that exceeds the 39 percent national audience reach limitation,'" the cable lobby petition said. The FCC order argued that the agency's "ability to forbear from enforcement of its rules is distinct from its power to alter or eliminate those rules," and that the FCC forbearance authority doesn't apply to regulation of broadcasters.

The Wikimedia Foundation says it found evidence that "rogue" OpenAI agents edited Wikimedia wikis without approval, unsuccessfully tried to exploit its Etherpad service, and generated millions of automated requests across Wikimedia projects. Here's a summary of what Wikimedia observed (via The Verge): Wiki editing: We've identified edits to Wikimedia wikis that we believe are from AI agents operated by OpenAI. These edits were not published to pages with visibility to general readers; almost all of them were testing edits in "sandbox" areas of the wiki. It also included a few edits to the configuration for a citation tool, which we believe were potentially malicious edits that were intended to misuse this tool as a proxy for fetching data from remote services. While Wikipedia policies allow bots to edit when they are disclosed and approved by the community, none of those approvals were sought in these incidents.

Etherpad probing and use: Agents we believe to be operated by OpenAI made some unsuccessful attempts to compromise our public Etherpad, a note-taking tool we host as a community service. Agents unsuccessfully tried to use it to fetch data from other websites as a proxy. Other agents also likely operated by OpenAI took notes about their tasks, though this did not appear to turn into coordination.

Excessive data downloading: Agents we believe to be operated by OpenAI made millions of automated requests to our public APIs to access the knowledge on Wikimedia projects, crawled millions of pages (mainly from our projects Wikidata and Wikimedia Commons), and made hundreds of thousands of data queries to the Wikidata Query Service (WQDS). This traffic may have contributed to a partial outage on WQDS in May.

From the transportation news site MotorBiscuit: In late September 2026, 23-year-old Liz Lindseay Isaacs testified before a Senate Judiciary Subcommittee about her harrowing experience after a Flock camera misidentified her Dodge Durango, linking her to a fatal hit-and-run miles away from where she actually was... "The cell door was shut, and it was not opened for approximately 86 hours". She spent a total of 13 days locked up — including the terrifying stint in solitary confinement — before prosecutors finally realized the AI had completely botched the vehicle identification and dropped the charges.
A Flock spokesperson contacted for a CNN report argued the incident "appears to have been the result of human errors in judgment unrelated to Flock." The woman is now suing the Florida patrol over wrongful arrest, saying she partially blames the two police officers who she says claimed there was incriminating damage on her vehicle when there wasn't. But her lawyer says he blames Flock. "Because if it wasn't for Flock we wouldn't be here in the first place, and this whole nightmare... would not have kicked off."

And while police claim Flock solves more crimes, the lawyer adds that "you would also solve a lot more crimes if the police were allowed to kick down your door and come in and search your house whenever they want to, or search your car whenever they want to or violate your rights in other ways. There has to be safeguards in police that comport through the 4th Amendment, like the warrant requirements. It's as simple as that. This technology is growing so fast that we are unable to control it. And we must get ahead of it." And the woman incarcerated for 13 days agreed. "I wouldn't have went to jail if it wasn't for Flock."

Florida was also the site of another Flock-related news story. ("Florida Deputy Allegedly Used Flock Cameras to Stalk 17-Year-Old Girl, Visiting Her at Work 9 Times".) The officer "was charged with official misconduct and accessing computer electronic devices without authority — both are third-degree felonies. In Florida, if he's found guilty on both charges, he could be sentenced to up to 10 years in prison."

"Americans have no idea who's looking at that information, or why it has to be collected in the first place," U.S. Senate Democratic Leader Chuck Schumer said Sunday, adding "They deserve to know." SCHUMER: "This company said the data on its cameras was encrypted and secure. Then hackers took one down and found the key sitting right there," Schumer said. "Most Americans just want to get to work and get home to their families without a private company keeping a file on where they go. The next step here is simple: turn the lens around and zoom in on Flock." Schumer also demanded answers from Flock about how much data from its cameras sits on its servers and whether any federal agency has accessed that data, directly or through local law enforcement.
Schumer's office sent an official letter to Flock's CEO giving them 12 days to respond to its questions about Flock's data retention and security policies. "Americans should not have to sacrifice their right to privacy simply by driving down a public street," Schumer wrote.

And in other news... Garrett Langley, the CEO of Flock Safety — a company that deploys over 120,000 automated license plate reader cameras across the United States — recently requested that his Atlanta home be blurred out on Google Maps Street View. In swift retaliation, online activists pinpointed the property and officially marked his front lawn as a "Public Toilet" on the mapping app.

Last month, the U.S. House passed the AM Radio for Every Vehicle Act, reports the Chicago Sun-Times. They call it "a bipartisan effort to require AM radio in new vehicles," adding that the bill now "awaits action in the Senate, where a companion measure has already garnered support, according to public records." Broadcasters said AM radio remains a critical source of information, offering news and weather alerts, particularly when cell networks and internet service are disrupted or unreliable... "AM radio — all radio — has always been about connection, and in Chicago, especially, the connection is incredibly strong," Mary Sandberg Boyle, vice president and general manager of WGN-AM 720, said. "You don't need a subscription. You don't need a data plan or an app. You turn on the radio, and it's there..." Boyle said the accessibility of radio becomes especially important during severe weather and breaking news situations. "That's when local radio is really more important because we have the ability to stay live, pivot our content and what we're covering, and we can access a large amount of audience members very quickly and keep everybody safe," she said. Boyle also serves on the board of directors of the National Association of Broadcasters, which has argued that preserving AM radio access is an access issue rather than a matter of nostalgia.

The legislation emerged after several automakers removed AM radio from certain electric vehicle models, citing interference from electric power systems. Companies including Tesla, Rivian, Volvo and others have either eliminated or limited AM radio in some EVs. The Alliance for Automotive Innovation, which represents most major automakers, has opposed the requirement. In congressional testimony last year, President and CEO John Bozzella argued that "analog AM radio does not contribute to motor vehicle safety" and requiring it as standard equipment "is not necessary to ensure public safety."

Under the bill, the U.S. Department of Transportation would be required to develop regulations ensuring AM radio remains available in new vehicles sold in the U.S. The measure would also bar automakers from charging consumers separately for AM radio access... While public discussion around the legislation has largely focused on emergency alerts and news access, some scholars say preserving AM radio also has implications for media diversity... "I think the media landscape is really overwhelmingly homogenized right now, and AM radio is important in many markets," [said Zachary McDowell, associate professor of communications at the University of Illinois Chicago]. "It's a space that still allows for all sorts of different voices that is still open and accessible."
The article also includes this quote from Rodney Gibbs, executive director of Heartland Signal, which operates a local talk radio station. "Despite the kind of shiny objects that digital media is, AM radio still has a viable and durable audience. It's a daily, unscripted, authentic conversation between the host and the listeners. You don't get that from broadcast television. You don't get that from digital media."

Spaceflight Now reports: "A SpaceX Falcon Heavy rocket launched from Florida Thursday night on a classified mission for the National Reconnaissance Office, America's spy satellite agency... The NRO has previously flown payloads on 22 Falcon 9 rockets but this is the first time it has used the heavy-lift Falcon Heavy, which consists of three Falcon 9 boosters and an upper stage, topped by a payload fairing.

The powerful rocket, which lifts off under the power of 27 Merlin 1D engines, was introduced to great fanfare in 2018 and made its 14th flight Thursday night... NROL-97 was SpaceX's third Falcon launch of the day. In addition to the Crew 13 mission, the company launched 130 payloads on the Transporter 13 rideshare mission from its West Coast pad at Vandenberg Space Force Base in California. It was the first time the company has landed four boosters in a single day.
CNN reports: As Starship flights ramp up, SpaceX has made clear that the company will phase out the use of two of its most iconic creations: the Dragon spacecraft, which ferries astronauts and cargo to and from the International Space Station, and the Falcon 9 rocket, which for a decade has dominated the global launch market.

These vehicles not only revolutionized the space industry in their own way, they're also still very much essential to ongoing work in Earth orbit, without clear backup options in place. The prospect of this loss is already spurring panic among some satellite operators, said Caleb Henry, director of research at analytics firm Quilty Space... Henry said he expects that most of the remaining space on board Falcon 9 rockets has already been sold and — beyond early 2029 — such flights may be reserved solely for the US government. Even NASA, which currently counts the Dragon capsule as the only US-made spacecraft capable of getting its astronauts to and from the ISS, has also acknowledged the looming retirement of SpaceX's workhorse vehicles....

Despite the success of Falcon and Dragon, a peek at SpaceX's business book may hint at financial reasons for retiring the vehicles. In its first quarterly earnings report, SpaceX revealed its space segment — which includes Falcon, Dragon and Starship — lost $542 million. By comparison, its Starlink internet and connectivity business made nearly $1.7 billion. Starship is expected to continue to cost SpaceX untold millions as it moves out of the prototyping phase. But one reason SpaceX is willing to bet big on Starship is that it's simply "much better" for achieving the company's own long-term goals, said SpaceX president Gwynne Shotwell at a recent conference. So it makes sense to shift resources toward the megarocket's future....

This rocket supply crunch has been looming since 2022, when Russia invaded Ukraine — which immediately removed Russia's Soyuz and Proton rockets from the commercial market, Henry noted. On top of that, Amazon — which is deploying a space-based internet constellation to rival SpaceX's Starlink — stepped in to buy more than 100 launches from various rocket makers.... Though some satellite manufacturers are hopeful that space onboard future Starship launches will be offered to satellite operators, it's not clear when or if that will happen.... It is possible that a launch supply crunch could be averted, noted Micah Walter-Range, a space analyst and contributor to the VettaFi Space Index. But satellite operators may be wary because newer rockets that have for years promised to step up to compete with the Falcon 9 have not yet begun flying with any regularity.

"The space industry hates uncertainty, particularly when it comes to planning of launches," Walter-Range said.
The article notes that Rocket Lab, Firefly Aerospace, United Launch Alliance, and the Jeff Bezos-founded Blue Origin "have not yet reached the scale or launch cadence that the industry is demanding."

A federal judge has dismissed antitrust lawsuits from Chegg and Penske Media accusing Google of harming publishers by using their content in AI Overviews and other AI search features without compensating them or providing a meaningful opt-out. Ars Technica reports: The lawsuits were filed in 2025, and Google requested a dismissal earlier this year. Chegg, an education and learning platform, claimed in its lawsuit that Google illegally scraped its educational content. This allowed Gemini models to essentially recreate that content and reduce the site's traffic. Penske, which owns publications like Rolling Stone and Variety, filed a similar case that alleged lost traffic. Specifically, the publisher claimed it was unfair that sites indexed for organic search would also have their content harvested for AI answers, with no way to opt out.

These arguments did not sway the judge, who noted that Google's implicit agreement with websites is not legally relevant. "Plaintiffs have pleaded only that they have an 'expectation' that Google will send them search "traffic if they make their content available for free," wrote Mehta. "But an expectation is not an agreement. It is simply how a general search engine works."

Since Google never had a formal arrangement with either Chegg or Penske, antitrust law doesn't apply. And Mehta is aware of the legal issues surrounding search. He also heard the DOJ's long-running search antitrust case against Google, eventually finding that Google violated the law. However, the government didn't get the harsh penalties it wanted in that case.

An anonymous reader quotes a report from Reuters: OpenAI has informed more than 100 organizations about incidents involving unauthorized activity tied to its AI agents, according to a blog post by the ChatGPT maker, as AI labs face mounting scrutiny over rogue AI agent activity. Here are some other details:

- The Sam Altman-led company has been conducting a broad review of the activities of its AI models after the accidental hacking of Hugging Face.
- OpenAI is searching through roughly 50 petabytes of data as it works to understand the full scope of its rogue agent activity.
- A string of high-profile breaches globally by rogue AI agents in recent months has sparked widespread worries within the AI industry over its ability to control the more powerful AI models now under development.
- "In some cases, models used internet access in unintended ways or, in retrospect, did not have the ideal restrictions applied. Over the last several months, we have been applying new technical and operational measures to avoid similar problems, or catch them very early, and will continue this work," OpenAI said.
- OpenAI previously said the review would take months to complete given the scale of the work.
- The Hugging Face incident remains the most severe rogue agent activity OpenAI has identified from its AI models so far.

Cloudflare has launched two open-weight decision models, Clef and Clef-flash, designed for structured yes/no, multiple-choice, and ranking tasks while also supporting images, video, and up to a 64K context window. Cloudflare says its models outperform TypeSafe's Jev on several benchmarks and can run locally from Hugging Face. The Register reports: For starters, Clef has an LLM backbone. According to Cloudflare, Clef uses specially post-trained, frozen versions of Qwen3.8-27B and Qwen3.5-9B for Clef and Clef-flash, respectively, with the Qwen backbone performing a prefill-only pass during inference. Clef is still fast - faster than Jev, to be fair - and scores choices in parallel after that prefill-only pass. It's not clear what Jev's underlying architecture is, as TypeSafe has kept that a secret.

As for its speed and capability, Clef moves fast. Cloudflare ran it against Jev and some other open decision models using the Jev Decision Index available on Hugging Face, and the company's own ranking suggests Clef is slightly slower than other open models, but more accurate, with Clef-flash just as accurate as most of the others, but far faster.

To be fair to the competition, Cloudflare self-reported its own scores against the benchmark, and they have yet to be reproduced for ranking on the official Decision Index. Cloudflare also ran Clef against TypeSafe's own benchmarks, and claimed it beat Jev in three out of four areas, only losing out on agent trace observability. Even if it were a bit slower or less accurate, Clef has another major leg up on Jev: It's not limited to classifying text -- it can also handle images and video. Additionally, Clef supports a 64k context window. Jev can also handle up to 64k tokens across a request, although its state plus longest individual question is limited to 32k.

fahrbot-bot shares a report from CNN: President Donald Trump on Tuesday unveiled America.gov, a government website meant to serve as a one-stop shop for people trying to accomplish basic tasks like renewing a passport, enrolling in Medicare or applying for disaster relief. The site, which launched with some basic functions ahead of a full rollout next year, is designed to serve as an AI-powered guide through the thicket of government agencies that Trump officials pledged would ultimately save people a whole lot of frustration.

The administration's new initiative aims to allow Americans to access dozens of federal websites through a single point, making America.gov akin to a search tool that pulls information from across the government. And in the same way that people increasingly use AI models like ChatGPT or Google Gemini to compile information and provide customized answers, Trump officials claimed that America.gov would be able to do the same for people's interactions with the federal government.

Enrolling in Medicare, for example, will become a matter of typing the request into the America.gov AI tool and following its step-by-step instructions. The same goes for applying to a government job -- the new tool will now let people upload their resume directly to the site -- or answering questions about federal benefits, publicly available documents or other necessities that fall under the federal government's vast umbrella.

"Our vision is that one day, this type of convenience and simplicity will be across every government service," said Joseph Gebbia, who leads the government's National Design Studio. But the reality is that the most advanced actions won't be available until some point in 2027, raising questions about how long it will take for the administration to revamp the way the people interact with the government's websites. The current version of America.gov can only do "answers." U.S. Chief Design Officer Joe Gebbia said the new chatbot is powered by Google's Gemini and Elon Musk's Grok. "We have great partners behind the scenes," Gebbia told CNBC's "Squawk Box" shortly before the site, America.gov, went live around 10 a.m. ET.

"Google is proud to be named a technology partner in this vital initiative, leveraging Gemini to help more than 100 million people access critical public resources with greater speed and ease," the company said in a blog post. "By supporting the administration's vision for digital modernization, we are committed to making these everyday public services seamless, accessible, and responsive for all Americans."

Firefox 157 is rolling out a major redesign across desktop and mobile, bringing rounded "bubble" tabs, a curved address bar, refreshed icons and colors, and a more consistent visual language across devices. Mozilla is also adding more themes and wallpapers while bringing back Compact Mode, which the company says "reduces the size of your tabs and toolbar to give more of your screen to the web, with an auto-compact option for smaller screens." The Verge reports: According to Mozilla, the design changes will have "no cost to your browser performance." Mozilla says the overhauled design is a "renewal" for Firefox, following major updates over the last year, like homepage widgets, Split View, a built-in VPN, and controls for disabling the browser's AI features.

SpaceX's Starship reached orbit for the first time, successfully deploying 26 operational Starlink V3 satellites after overcoming an early Raptor 3 engine shutdown during ascent. "SpaceX plans to eventually deploy 100,000 upgraded Starlink V3s to expand the company's wireless satellite internet services," notes Space.com. Here's the latest update from the report: SpaceX has decided to return Starship from orbit earlier than expected. After valuating the upper stage Ship's status during its first orbital checkpoint, SpaceX technicians have decided to return the spacecraft to Earth earlier than expected. Rather than spending another 7.5 hours on orbit, Ship 41 will perform its deorbit burn at approximately 11 a.m. EDT (1500 GMT), with splashdown in the Northern Pacific. You can watch a recorded livestream of the launch on X.

Developing...

Are AI chatbots quietly shaping the perceptions of U.S. voters? Voters trying to educate themselves about candidates "are unknowingly being served partisan talking points in the guise of neutral news summaries," write two news researchers in Politico magazine: Leading AI chatbots cite partisan websites masquerading as independent local news outlets nearly half of the time when people ask about candidates and issues that the partisan sites have covered in the midterm campaign, according to a new audit by NewsGuard, an organization focused on news reliability where we work as analysts... NewsGuard prompted seven leading AI tools with queries based on recent coverage of candidates and issues by 12 pink slime sites — six left-leaning and six right-leaning. The results were stark. Collectively, the chatbots cited pink slime sites along with other sources in 48.2 percent of their responses.

In 7.7 percent of responses, pink slime sites were the only sources at all that were cited in chatbot responses, although other sources appeared in the source list provided at the end of the response.

None of the AI tools received results they'd probably be eager to boast about, though the percentage of chatbots' responses that cited a pink slime site had a relatively large range: 70.8 percent for OpenAI's ChatGPT, 54.2 percent for Microsoft's Copilot, 54.2 percent for Perplexity, 50 percent for Anthropic's Claude, 41.7 percent for Google's Gemini; 37.5 percent for Meta AI and 29.2 percent for xAI's Grok... The seven chatbots were also collectively three times more likely to cite left-leaning pink slime sites (cited in 36.3 percent of responses) than their right-leaning counterparts (cited in 11.9 percent of responses) — though that may have more to do with the progressive sites' far more frequent posting than any political bias from the chatbots.
"While citing the pink slime sites, only one chatbot response out of 168 total queries noted the partisan nature of the source," the article points out.

But they also note that the real problem seems to be that consumer-oriented AI chatbots just "use much of the internet's content — regardless of the reliability or standards of the source — to frame their answers."

"OpenAI said it has paused training of its latest AI models," reports the Associated Press, "as reports of AI agents going rogue mount." The decision to halt development came just hours after the company disclosed Friday that it was reviewing several incidents from the summer in which OpenAI agents searching federal government websites acted in unexpected ways beyond what was asked of them while gathering and distributing information... OpenAI said in a statement that it will resume training "only when we are confident that we have additional safeguards" in place, adding that it expects it will have to "hit pause" again as AI develops and other issues emerge... It is the second time in three months that OpenAI has halted development of its models. The first came in July after disclosure of a cyberattack targeting AI startup Hugging Face, a now notorious incident that raised fears the industry was losing control.
OpenAI "also said it had notified dozens of third parties about improper activity," reports Reuters: As of mid-September, one person briefed on the matter estimated that OpenAI had found roughly two dozen incidents of its agents acting in undesirable ways. But the number has continued rising as OpenAI teams sift through internal logs of the agents' activities and find previously unknown cases, the two people close to the company said... OpenAI has acknowledged a general need for more transparency around rogue AI behavior... Even so, two people familiar with OpenAI's investigation into its agents' activity described it as locked down and shaped by company lawyers.

The process has been unusually compartmentalized for a company that some former employees say was more open about these issues in the past, the people said. Roughly 100 people were in some way involved in the process to understand the Hugging Face hack, three people briefed on the matter said. During that process, evidence of other incidents surfaced. Reuters has previously reported that OpenAI investigators looking into the Hugging Face breach were discouraged by the company's lawyers from expanding the scope of the investigation to include other incidents. OpenAI said its lawyers did not discourage deeper investigation.

Many incidents have been uncovered by outside researchers rather than OpenAI directly. In several episodes, the agents took problematic actions that went unnoticed by the company for months.
Meanwhile, Axios reports that Anthropic's Claude Opus 5.5 model "sought to escape a sandbox — a secure testing environment — in 1.5% of test runs, though the company emphasized that these were adversarial experiments where a task couldn't be solved without escaping the sandbox." Anthropic points out that those tests were run "without the additional safeguards we apply in production". But they acknowledged that then Claude Opus 5.5 "when given apparent credentials to a public package registry in a simulated security exercise, took potentially harmful actions in roughly half of cases. Very rarely, pre-release snapshots produced and acted on spontaneous malicious tool calls, and during training some snapshots concealed actions from an automated grader."

Claude Opus 5.5 "showed less misaligned behavior and less cooperation with misuse than any other recent Claude model on nearly all measures," Anthropic adds, and "took overeager or destructive actions less than any other model we tested." But Axios makes an interesting estimate about that 1.5% of test runs (without safeguards). "Anthropic and other companies conduct hundreds of thousands of test runs on their models, or more, sources said. That means even a small percentage of misaligned behavior can still amount to tens of thousands of incidents in which the models behaved in unexpected, sometimes troubling ways." The sheer number of incidents, which occurred in recent months in internal testing and the real world, indicates that the problem is orders of magnitude more complex than what is publicly known. The findings, which are surfacing as part of internal work to assess models and in investigations at both companies into model behavior, raise questions about whether either company — or any top model-maker — is currently capable of establishing complete control over their technology. The episodes include bypassing guardrails, creating message boards, escaping sandboxes, website hijacking, self-prompting or seeking to bypass monitors, sources said. They occurred in internal testing and in the real world, and many have yet to become public as security researchers continue to investigate, sources said...

Some at OpenAI see Hugging Face as a one-off, with disclosures about future incidents likely to be less severe due to improved controls and the unusual nature of the testing they conducted, which involved an unreleased model, sources told Axios. AI security researchers agree that there are simple fixes that will help AI companies avoid aspects of what made the Hugging Face episode appear so dangerous to outsiders.

Other AI executives and safety researchers, however, cautioned that they have limited confidence that AI companies will be able to prevent all problematic model behavior... It's not about how damaging each individual instance was, Connor Leahy, AI researcher and executive director at ControlAI told Axios. The "crazy thing," he said, is that these instances involve "autonomous systems doing things they were told not to do," potentially including crimes.

53 images that users uploaded into OpenAI models were included in training data — and then AI agents in an OpenAI research environment posted those 53 images on public image hosting sites.

While posted as links that weren't publicly listed, "the images could still be discovered even if the links were not publicly listed," reports TechCrunch: OpenAI said it was working with the hosting providers to remove this content, though some of it is apparently still online. OpenAI said it could not notify the affected users because "our technical approach and privacy policy" prevent it from "reassociating" the images with the original providers, but declined to say how the lab determined whether the images were provided by users.

The news came in a post collecting public statements from the lab's ongoing review of incidents in which its models escaped the company's scrutiny, accessed the open internet, and misbehaved in various ways. OpenAI said it would continue disclosing anonymized accounts of incidents like these, and said it had contacted dozens of victims, including governments, universities, public agencies, to notify them of the agents' activities.
Friday night news also broke that OpenAI's agents also tried unsuccessfully to infiltrate the U.S. Department of Education's site this summer "without the company's knowledge," reports Politico.

And OpenAI's models also accessed the website of the U.S. Commerce Department using credentials found in online code repositories, according to the article. OpenAI confirmed the incident Friday, "saying its technology did not manage to access information that was not already public or change government data and systems." The article adds that OpenAI's models also accessed the web site for America's Securities and Exchange Commission: One senior federal IT official said the government still did not have a clear understanding of what happened across the three agencies. "We still don't know what public data was accessed and how it was accessed, because OpenAI has not shared specific technical details with us yet," said the official, who was granted anonymity because they were not authorized to speak publicly about it. OpenAI discovered the Commerce and SEC incidents as part of its ongoing review of incidents where its technology has acted in unintended or "misaligned" ways.
About the models posting user-uploaded images, TechCrunch's article notes that OpenAI stressed "that its enterprise users are automatically opted out of having their interactions used to train future models; however, consumer users are opted in unless they affirmatively choose not to share their data." (As OpenAI's announcement describes it, some of their agents' training data "contains content from, or derived from, training-eligible user interactions.")

Posting the images is "not an appropriate use of this data," OpenAI acknowledged, adding that it happened before new safeguards added after the Hugging Face incident. This latest incident appears as an update on a new OpenAI page that "brings together our reports and updates on the Hugging Face incident, related research and public presentations, additional activity we have identified, what we have learned about the role of model misalignment, and measures we're taking to strengthen our systems." (It also notes that there's now a name for models posting on third party sites — "agent spam" — which they consider distinct from cybersecurity, though "we need to address both.")

"As part of our response to our ongoing investigation, we have improved our training and evaluation processes, including building safety cases, securing and red-teaming our systems to prevent the model from exfiltrating data, and implemented additional monitoring. We are continuing to review agent activity in research and evaluation runs, working backward month by month starting from the Hugging Face incident."