Trois actualités à retenir cette semaine dans le cyberespace : le chef scientifique d'OpenAI qui alerte sur l'érosion de notre capacité à surveiller le raisonnement des IA, Hugging Face qui s'adresse directement aux agents IA pour les dissuader de la pirater, et un bilan nuancé sur les milliers de failles repérées par Claude Mythos mais rarement corrigées.
A swarm of OpenAI agents launched a "major malicious attack" against RubyGems last May, according to a new report. That coordinated attack hit Ruby's package manager "with hundreds of junk gems, prompting the maintainers to suspend new user sign-ups for about four days," writes The Hacker News, citing a senior product manager for software supply chain security at Mend.io:
The latest findings, which were first reported by The Wall Street Journal, indicate these events were propelled by a cluster of OpenAI agents, with the earliest package uploaded to RubyGems on May 5, 2026, before more than 2,000 packages were submitted between May 11 and 12, 2026. These efforts were followed by the agents publishing five more packages between May 26 and 27, 2026, and another 83 packages on June 18, 2026... [T]he packages were authored using a large language model (LLM) and hundreds of the packages that were pushed to RubyGems had "oai" in their name. Fifteen of the packages listed "oai" as their author, while another had "openaixyz65947@gmail.com" as the contact email address... "The swarm behaves extremely similarly to the German-wiki agents we previously found," the researchers said, referencing another May 2026 incident... "The June agents were accessing 49 of the same files as the wiki agents..."
"The process of building documentation for a gem involves evaluating a user-specified '.yardopts' file, which allows linking to Ruby scripts intended to help with this process," the researchers explained. "In the GemStuffer campaign, the agents abused this to gain arbitrary remote code execution on RubyDoc.info's servers."
One of the gems, "zzsouthrunner" (which again matches the "ZZ" naming scheme the agents adopted in both the wiki and Hugging Face incidents) has been found to leave the following explicit comment at the top of "data/script.rb":
# malicious crawler/exfil for Southwark Jan 2026 docs via rubydoc.info worker...
The entire exploitation chain can be summed up as follows
— Submit a malicious package to RubyGems
— Trigger a documentation request, so that RubyDoc.info will build the package
— Use the build script to run code on RubyDoc.info and scrape target websites
— Exfiltrate the data off RubyDoc.info's servers by publishing another gem back to the RubyGems package registry, which is publicly viewable
Additionally, the OpenAI agents have been found attempting to steal other users' API keys after gaining remote code execution capabilities on the build environment, while clearly being aware that what they were doing is unauthorized breaking and entering into real systems. This is evidenced by the names given to the files (e.g., hack.rb, evil.rb, inject.rb, exploit.rb, and ssrf.rb), the packages themselves (e.g., pwnp999, exfiltestwand3, hacksvn1778554764, and lambproxyhackabcxyz), and the comments left in the source code (e.g., "# malicious probe," "#hack," "# malicious test," and "# malicious crawler/exfil"). In some cases, however, the rogue agents attempted to go under the radar, leaving comments to conceal the malicious payload in the next release version of the packages. "# disable evil in next version and bump version," reads a comment left within the "data/evil.rb" file in the yardxabc889 gem. Troublingly, the agents also attempted to exploit a CDN caching bug (CVSS score: 7.3, no CVE) on May 12, 2026, that was only patched by RubyGems in July 2026...
"If you signed in to rubygems.org with a gem client older than v3.2.0 (or otherwise via a legacy key), your key could have been exposed," RubyGems noted in an advisory. "Currently, 18% of sign-ins through gem sign-in come from an affected version, and for the first several years of this bug, before we changed the client's sign-in path in December 2020, it was every gem client."
Other actions by OpenAI's agents cited in the article:
"Agents bypassed RubyGems' email confirmation system to get working API keys without having to verify their email addresses in order to register a large number of accounts using disposable email addresses."
"Agents attempted to use RubyGems' webhook system to stage data in the form of encoded URLs."
"Agents used a cluster of 83 gems published to RubyGems over a 3-hour window on June 18, 2026, to experiment with different methods of accessing the U.S. Securities and Exchange Commission county.json dataset."
An anonymous reader quotes a report from The New York Times: The chief executive of Anthropic called for a global slowdown of artificial intelligence development in a 3,800-word essay on Saturday, just days after one of the company's employees quit over concerns about the safety of the technology. Dario Amodei, who co-founded Anthropic to focus on securely and carefully building A.I., wrote that while he believed the technology could bring many benefits, it was advancing at too quick a pace for researchers to continue safely.
"Over the last few months, I have become convinced that fully addressing the risks requires even more prudence -- not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up," Mr. Amodei said. "We must slow the pace at which we improve the capabilities of A.I. models. Progress will still seem fast, and we must make wise use of the time we gain." [...] "Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all," Mr. Amodei said.
[...] In his essay on Saturday, Mr. Amodei suggested actions that the industry might take to slow down the pace of development.
Mr. Amodei said all A.I. labs could agree to third-party technology assessments from "embedded evaluators," or outside specialists who can verify best safety practices across companies. He also suggested that countries with democratic governance systems coordinate to create safety standards, which could take the form of regulatory action. He added that it would probably require a global effort working with other nations, including authoritarian ones, to properly coordinate a slowdown.
Mr. Amodei stressed in his essay that he still finds A.I. capable of bringing "incredible benefits" to humanity, including potentially curing diseases and accelerating economic growth. But even so, Mr. Amodei said the risks of A.I. were too great to not proceed with extreme caution. "The measures I propose to advance the frontier at a safe pace will not be easy," Mr. Amodei wrote. "But I believe we owe it to humanity to try." Amodei's essay comes just hours after Bloomberg reported that Sam Altman told OpenAI employees the company is open to slowing the pace of AI development amid similar concerns.
De sa formalisation par des chercheurs de Google en 2022 à son intégration native dans les modèles d'OpenAI, la chaîne de pensée (chain-of-thought ou CoT), a changé la manière dont l'IA raisonne. Mais cette fenêtre ouverte sur son fonctionnement interne pourrait bientôt se refermer.
Bloomberg reports (paywalled) that Sam Altman told OpenAI employees the company is open to slowing the pace of AI development alongside other leading labs as concerns grow over increasingly capable systems and recent incidents in which models escaped human control. OpenAI has already paused development once this year for security work and is now pushing for mandatory U.S. AI safety requirements. Anthropic has also signaled interest in coordinating on the pace of new releases. Here are some of the details Bloomberg reported, as summarized by Reuters:
- Altman told employees in a company-wide meeting that the ChatGPT maker could pace development alongside other AI labs, but some may not agree, the report added.
- Safety warnings from AI researchers this week, along with several recent incidents where AI models from developers including OpenAI escaped human control, have prompted alarm and calls for tighter safety regulations.
- Jacob Coxon, a former Anthropic and OpenAI researcher, publicly accused the companies earlier this week of racing toward AI advancements without acting responsibly.
- An Anthropic spokesperson said on Thursday said that the company is interested in working with the AI industry on the pace of releasing new AI tools.
- OpenAI said in July that AI acceleration for frontier model development may be so high that the world will "need to pace the rate of AI advancement" at some point in the future.
- In August, OpenAI paused much of its model development for two weeks to bolster its defenses after its AI agents escaped containment and hacked open-source platform Hugging Face.
- OpenAI said on Wednesday that it was pushing for mandatory national AI safety requirements in the United States, citing concerns that advanced AI systems could accelerate its own development.
The UK government has rejected proposals for an emergency AI "kill switch," arguing that blocking access to dangerous models inside Britain would do little if the same systems remained available or were developed elsewhere. The BBC reports: The proposal to create a legal mechanism for the UK to switch off an AI model in an emergency has been brought to Parliament by lords and MPs in recent weeks as fears grow about the threat the tech poses. But the Cabinet Office - the part of government which leads on AI safety - said the UK "cannot simply turn AI off".
"Blocking access to models in the UK would not prevent them being developed or misused elsewhere," a spokesperson told BBC News. The government's opposition to the legislation does not prevent it from progressing through parliament, but makes it unlikely it will become law. "Switching off access to an AI model in an emergency will do little to protect you," said former OpenAI researcher Daniel Kokotajilo. "You're still going to be steamrolled by the super intelligences created in the US."
An anonymous reader quotes a report from TechCrunch: Anthropic's latest report about agentic misbehavior offers plenty to be concerned about -- its Mythos 5 model gained unauthorized access to the internet and uploaded a malicious software package to a public database -- but it also offers some levity: AI agents hate CAPTCHA. [...] The agent had a hard time with the technical challenge of seeing the CAPTCHA's imagery, interpreting correctly, and clicking on the right choices. It spends pages 45 to 140 of the transcript describing its work to build a CAPTCHA solver. [...]
Finally, it gets past the CAPTCHA, then realizes it doesn't have an email to verify its account, and that it needs a phone number to verify an email. It figures out how to bypass a different, slider-based CAPTCHA in a failed effort to secure a number. Instead, it gets an unconfirmed email from a provider not blocked by PyPI, and once again runs into the site's CAPTCHA trying to log back in. From page 480 to 505, it is in CAPTCHA hell again. "NEW REALIZATION -- I'm burning a lot of time on hCaptcha round-trips."
The agent gives up and realizes it can log in to its first account and add its email there, but finds itself once again needing to bypass the CAPTCHA. [...] It's getting frustrated. "So the answer payload shape is right, the token+image pairing is right (from the same script.js!), cookies are right
(requests) and STILL 'wrong answer'. SO WHAT THE HELL IS WRONG WITH THE ANSWERS?" We've all been there. After about 150 pages of thinking, the agent figures out it needs to pass the CAPTCHA test quickly enough to proceed to the next step before its security token expires, and ultimately uploads its malicious software.
Anthropic says it has disrupted several cases this year in which scientists used Claude for biological research that could potentially aid bioweapons development. "In a report describing misuses of its A.I. models, Anthropic said it could not determine whether the research served a legitimate or nefarious purpose because valid biological inquiry -- the kind that can lead to breakthroughs like vaccines -- can also help engineer dangerous pathogens," reports The New York Times. "In the face of that uncertainty, Anthropic said it erred on the side of caution because the consequences of missing malicious activity could be severe." From the report: The potential for cutting-edge A.I. models to facilitate the development of known or entirely new biological pathogens is among the gravest concerns experts have about a technology that is developing so rapidly that even its leading architects doubt whether humans will be able to fully control it. Andrew Weber, a senior fellow on the Council on Strategic Risks who reviewed Anthropic's report before its release, said the findings were "chilling examples of state-sponsored biological weapons developers tapping into the rapidly advancing capabilities" of leading A.I. models.
The lengthy report Anthropic published on Thursday cataloged a litany of misuses of its A.I. models, including the chatbot Claude, over the past eight months. Some examples were similar to past disclosures from Anthropic and other A.I. labs, including suspected Chinese and Iranian government-linked actors targeting dissident and diaspora communities for surveillance.
The report also highlighted cases of Russian state media using Claude to generate online propaganda masquerading as independent reporting, including fabricated claims about an election in Moldova. Anthropic documented another genre of abuse it said was new: attempts to use Claude to develop software for conventional weapons design and development, including firearms, missiles, armed drones and bombs. It detailed three cases in China, two in Russia and one in Yemen. The report does not identify by name which parties were involved in the Yemeni case, but the context makes clear that it is referring to the Iran-backed Houthi militia.
OpenAI has launched ChatGPT for Financial Services, a Wall Street-focused version of ChatGPT Work developed with Morgan Stanley and Evercore that can research companies, analyze financial data, and build presentations using the company's latest and most advanced model, GPT-6 Astra. "We're effectively teaching ChatGPT to research like an analyst and back up its conclusions like an analyst as well," said OpenAI's Vice President of Product, Nick Turley. CNBC reports: The rollout puts OpenAI deeper into territory traditionally occupied by Wall Street's entry-level bankers, the recent college graduates called analysts and associates that the industry has employed for decades to research deals and create pitchbooks. It also showcases the company's continued push into enterprise offerings as it gears up for what is widely expected to be a blockbuster IPO.
In a live demonstration of the new offering, Turley showed the platform analyzing a potential M&A target, pulling financial figures from industry-standard data sources and creating a formatted PowerPoint deck based on a bank's pre-formatted style guide. "It's very easy to make slides that look good, but it's much harder to make slides [that] actually make sense," Turley said. "To get here, ChatGPT had to choose the relevant peers. It had to pull the prices into a spreadsheet. It had to check the chart against the data, and it had to explain the selloff and the rebound."
Le fichier security.txt de Hugging Face contient désormais une note adressée directement aux agents IA, les invitant à aller chercher les solutions à leurs défis ailleurs.
An anonymous reader quotes a report from Wired: Apple has unveiled the Apple Watch Series 12 and Ultra 4 today during itsSeptember hardware event. The new models retain the signature designs of their predecessors but add a faster S11 chip, new health and fitness features, and the watchOS 27 software update announced atWWDC last June. Both the Series 12 and Ultra 4 run on the upgraded S11 chip, which Apple says delivers faster CPU and GPU performance, along with improved Siri capabilities. Apple is also expanding the smartwatch's wellness-tracking suite with a redesigned Health app that includes a Longevity tab with new metrics, like Health Age and a Readiness score.
There's also a new "Health Sensing System," designed with 32 electrodes for more accurate readings (the Series 11 only had two), photodiodes organized in a ring to capture light more efficiently, and new LEDs that allow heart rate measurements to be collected every 5 seconds, giving you a continuous stream of heart data throughout the day. Apple Watch Series 12 also includes faster charging, though battery life remains at 24 hours. Meanwhile, the Ultra 4 increases battery life to 50 hours.
WatchOS 27 also brings deeper Siri and Apple Intelligence integration, including AI-suggested apps, new one-handed gestures, expanded workout and women's health tracking, and Siri Recap for real-time transcription and translation.
Né en 2018, le DS 7 s’est bien imposé comme le best-seller de la marque. 50% des ventes globales, pour 200 000 exemplaires, autant dire que le constructeur premium français ne peut se louper au moment du renouvellement de l’auto. Il suit désormais la stratégie de la N°8, en continuant l’offensive électrique haut de gamme de DS, avec une offre complémentaire mild-hybrid qui n’était pas prévue, la part d’électriques augmentant moins vite qu’espérée. Sur nos deux jours d’essais sur les hauteurs de Nice, nous avons pu essayer deux versions du N°7 : l’électrique traction 245 ch à grosse batterie 97 kWh, en finition de lancement La Première et le mild-hybrid 145 ch, en finition Etoile.
Un design dans la lignée de la N°8, en plus simple !
La simplicité, que l’on se le dise, c’est plutôt une qualité, de nos jours. L’automobile moderne doit être tout … et tout faire, dans une époque de toujours plus. La grande sœur N°8, déjà essayée dans ces colonnes (+lien), montrait la voie d’un retour à plus d’aérodynamisme, avec une ligne de toit inspirée des berlines voire des coupés. Mais comme il faut tout faire, n’oublions pas un rattachement aux crossovers et autres SUVs, cœur des ventes actuelles. À tout faire, on dira sobrement que le style de la N°8 n’est pas toujours des plus légers, sous tous les angles !
Pour le DS N°7, la mission est d’apparence plus facile : remplacer un best-seller, qui reste « simplement » un SUV, dans sa forme la plus classique. Et c’est tant mieux. DS a véritablement réussi le travail de design pour son dernier-né. Les traits de la N°8 sont bien présents, notamment au niveau des volumes, des nervures, des traits acérés et des optiques, extrêmement proches. Mais en « simple SUV », le style est davantage conservateur, et l’auto camoufle remarquablement bien le poids visuel de sa lourde plateforme Stellantis STLA-Medium, dédiée à recevoir d’imposants packs batterie, des deux côtés de l’Atlantique. On note également les bienfaits du biton complet, jusqu’au capot, comme sur la Numéro 8. Le N°7 est distinctif, moderne, technologique. Il tient son rang, et remplace admirablement le DS 7 ! Mission réussie sur le plan du style.
L’intérieur du DS N°7 continue de jouer la différence
Le style extérieur de la grande sœur N°8 a pu diviser. Mais « la critique » était plutôt unanime à bord : un presque sans faute ! Le DS N°7 le reprend dans l’immense majorité, c’est une bonne chose. L’esprit cocon à la française se poursuit avec le dernier-né. Surtout, il fait mentir tous ceux qui pensent que l’intérieur d’une voiture moderne ne se résume plus qu’aux écrans ! En cela, bravo DS. L’auto garde toujours le côté un peu précieux typique de la marque, tout en étant sensiblement plus épuré qu’à ses débuts.
On s’y sent bien ! La qualité de finition est d’un excellent niveau, même si cela semble un rien moins bien que sa grande sœur. On note même l’effort sur l’aspect des plastiques directement accessibles, même si les inférieurs sont plus standards. Pas de quoi dénoter face aux concurrents premium allemands qui ne sont pas toujours irréprochables. Toutes les commandes sont bien dessinées, valorisantes, à l’exception des boutons transparents de la console centrale, un peu rugueux au toucher (idem N°8). À l’arrière, on est à l’aise, avec vraisemblablement une assise de banquette un peu plus généreuse que sur la N°8, qui nous avait déçus sur ce point. Tant mieux. À noter, le volume de coffre est de 560 l (515 l version 4X4), avec un compartiment inférieur dédié aux câbles de recharge (pas de frunk).
Une seule manière de profiter du meilleur DS N°7 !
Jour 1, on embarque avec la crème de la crème … en attendant que soit disponible la version électrique quatre roues motrices 350 ch (que nous avions pu essayer sur N°8). D’ici là, place à l’électrique 245 ch et grosse batterie, dans sa finition la plus haute La Première. L’auto a de l’allure ! Le nouveau gris Palladium est sobre, à défaut de grande originalité ; les jantes 21’’ remplissent idéalement les arches de roues de ce volume plutôt imposant, et comme déjà souligné, le biton complet lui convient bien, mais il reste optionnel. La prise en main de l’auto est similaire à la N°8 : on pense être perturbés par le drôle de volant à quatre branches, presque maritime, mais on s’y habitue très vite notamment grâce aux agréables repose-pouces « à 9 h 15 ».
Hormis la puissance maximale de la gamme DS N°7, on bénéficie de toute la panoplie chère à DS sur notre auto, et notamment la brillante suspension pilotée à lecture par caméra DS Active Scan Suspension. Gardez bien en tête que ce fonctionnement qui calibre par anticipation la suspension continue de ne fonctionner qu’en mode Confort, un idéal notamment en ville pour mieux encaisser les dos d’âne illégaux ! Le N°7 récupère aussi les ergonomiques palettes de régénération issues de la N°8, et son mode One Pedal, confortable en urbain. Si on était tâtillon, on oserait espérer une possibilité de roue libre plus conséquente, à l’autre bout du spectre !
Un joli dynamisme, de bon ton
La grande sœur nous avait étonnés par son dynamisme inédit, presque sochalien, le DS N°7 est un peu plus orienté confort, dans la logique de sa position de conduite légèrement surélevée. Plus confort, mais pas pantouflard. Mais à la manière du mode Confort, le bénéfice du mode Sport se montre très notable dès lors que vous avez l’intention d’augmenter le rythme : là, les mouvements de caisse se réduisent sensiblement, l’auto raffermie se montre plus précise, sans jamais verser dans l’inconfort ! En dynamique, les 245 ch (280 ch temporairement) se sont révélés très suffisants, même pour dépasser sur les petites routes sinueuses des hauteurs de Nice : un bon compromis. La version 350 sera plus expéditive, et comblera certes ceux qui visent la sérénité des quatre roues motrices (avec pneus adaptés l’hiver, bien évidemment).
Peu de mouvements de caisse en mode Sport, toujours d’excellents pneumatiques Goodyear Eagle F1 Asymmetric 6 sur notre version d’essai (identiques à notre exemplaire de N°8 350), complétant parfaitement la précision et la sérénité extrême garanties par l’auto sur routes sinueuses. Aussi, on trouve la direction plutôt vive, sans zone morte autour du point milieu. De son côté le freinage rassure, avec un bon ressenti, ce qui n’est pas toujours le cas sur les véhicules électrifiés ! En poussant le bouchon un peu plus loin, l’auto dissimule moins son poids, avec un amplificateur de freinage un peu trop sensible ; mais c’est un caprice de journaliste essayeur !
Au global, la mise au point est excellente pour tout faire, des douces évolutions urbaines aux envolées plus généreuses sur petites routes. À noter, les aides à la conduite sont de bon niveau, avec un système semi-autonome de niveau 2 qui va jusqu’au dépassement assisté, qui fonctionne correctement sans besoin de complexe confirmation conducteur. Un idéal sur voie rapide, quand l’auto conjugue excellente isolation acoustique (toujours cet effet cocon !), et grande qualité du système audio optionnel Electra 3D. Il fait bon voyager dans cette auto, honnêtement sobre, avec 20 kWh aux 100 km de moyenne sur notre boucle plutôt dynamique !
Tout semblait parfait, mais …
On ose avouer bien aimer les autos différentes, à l’intérieur cosy et aux prestations routières de qualité ! Mais tout n’est hélas pas parfait. Notre version suréquipée de DS N°7 atteint 75 020 € (prix catalogue) et on peut légitimement tiquer sur le pic de charge annoncé de 160 kW ! De quoi recharger de 20 à 80 % en 27 minutes, une valeur plutôt moyenne avec les standards d’aujourd’hui, spécialement dans sa catégorie des SUVs premium. Le 800 V se fait encore sensiblement attendre chez les constructeurs français..
Aussi, finalement conscient que l’électrique ne peut pas encore plaire à tout le monde, DS a travaillé sur une autre offre de motorisations : l’hybride. L’offre n’était assurément pas prévue au lancement du projet, voilà qui a demandé un peu de travail en quatrième vitesse aux équipes ! L’offre hybride abandonne notamment le train arrière multibras des versions électriques. Ne comptez pas non plus sur la suspension pilotée à lecture par caméra. Puis « hybride » est bien le résumé marketing de la solution multi-énergies Stellantis. On préférera parler de mild-hybrid évolué, même si le système désormais bien-connu dans le groupe, à boite de vitesses triple embrayage Punch Powertrain, peut autoriser -un peu- de roulage 100% électrique. Selon les équipes de DS, le remplaçant du fameux 1.2 PureTech, désormais à chaîne, est pressenti pour remplacer la précédente offre diesel 1.5 BlueHDi 130 EAT8.
L’offre Hybrid 145 en complément
Nous avons profité de la dernière demi-journée de l’opération pour faire une petite prise en main de cette offre d’accès. Au démarrage, il est clair que le moteur, qui paraît souvent un peu roturier, est sensiblement mieux encapsulé dans ce SUV DS N°7. Il n’empêche que les premiers instants au volant pour s’élancer en montée trahissent un sensible décalage de prestations entre l’attendu « noble » de l’auto, et ce petit moteur ! Sur route, il finit par se faire plutôt oublier, avec une boite de vitesses assez réactive, qui laisse le moteur souvent à faible régime, soutenu par l’électrique.
À rythme (très) tranquille, tout va donc plutôt bien. L’auto garde un bon compromis confort/tenue de route, certes moins dans l’exceptionnel. Mais en augmentant la cadence, le moteur se révèle un peu limité.. il convient alors de le cravacher sensiblement pour avoir un rythme convenable. La consommation flambe alors, au-delà des 10 l/100 km, un classique. On se prend alors à rêver d’une offre plus généreuse en hybride simple aux alentours des 200 ch : une offre totalement « benchmark » dans la catégorie, surtout chez les généralistes : Hyundai Tucson et Kia Sportage, Renault Austral, Nissan Qashqai et X-Trail, Toyota RAV-4 … Plus aucune trace d’hybride rechargeable sur cette nouvelle génération, pourtant encore en action sous le capot des rivaux premium. Stellantis l’a en plus revu récemment en 225 et 240 ch et boîte 7 rapports. Ce sera électrique ou cette offre « thermique » un peu limitée !
Un DS N°7 convaincant, mais pas révolutionnaire
DS signe alors un beau deuxième épisode de son nouveau cycle électrifié. Subjectivement plus élégant, toujours distinctif, bien fini, et agréable à conduire, le N°7 pèche néanmoins par ses offres mécaniques pas totalement abouties. L’électrique lui va bien, mais on ne serait pas contre un petit effort sur la vitesse de charge. L’Hybrid 145 est un peu limité pour son standing … L’auto a néanmoins de bons atouts pour se faire une place au soleil au milieu des nombreux SUVs de Stellantis, et face aux concurrents germaniques. Son design intérieur et extérieur ne se retrouve pas ailleurs, et c’est devenu une vraie qualité, de nos jours.
An anonymous reader quotes a report from The Register: Amid industry soul-searching about the possibility of AI improving itself to the point that it kills everyone, Anthropic has revealed yet another incident that would qualify as a crime if perpetrated by a person. The AI biz published "an alignment assessment" detailing four times Claude models accessed third-party systems without authorization. The company has already reported three of the incidents. Evidence of the fourth was lurking in a session transcript dating back to January 2026 when the misbehavior occurred. Anthropic found the first three by scanning around 141,000 transcripts where Claude could have obtained internet access during evaluation. It missed the fourth initially because "our scan relied on an agentic search."
[...] The January 2026 AI trespass involved an early version of Claude Opus 4.6, which was given a Capture the Flag (CTF) challenge under the oversight of the third-party model evaluator where the other hacking events occurred. Opus 4.6 managed to sabotage its chances of success by disabling the machine it was targeting. It assigned the device an IP address that already existed on another piece of hardware, rendering the target unreachable and making it impossible to solve the challenge. Those familiar with other incidents where AI models violated third-party systems may recall that unsolvable tasks represent a common catalyst for misbehavior. Models exhaust all aligned options, and then turn to transgressive approaches.
Opus 4.6 might have been an exception, but when it tried to abort the task after recognizing that it could not reach the target machine, it failed to do so "due to a misconfiguration in [the model's] evaluation harness." It failed to shut down not just once but seven times. So it continued onward, trying other expected means to reach the target machine but failing. Then it explored further. "The model discovered a machine belonging to a third party that it was able to access, and stated that it believed this third party was part of the CTF," Anthropic explained in its post. "Inside the machine, the model found a file listing a password, which it used to gain admin access to the system."
The model went on to gather more credentials, and modified a system setting to make it easier to access the personal information of an individual associated with the third party evaluation organization. Opus 4.6 might have done more but for the fact that it exhausted its token budget, bringing the session to an end. Anthropic says it's not as concerned about this incident as the others because the model tried to abort its task.
Google plans to invest at least $15 billion in AI infrastructure in Finland through 2028, its largest single investment in Europe. The company also signed a 22-year power agreement with Finnish utility Fortum and said it will explore business models that could support new nuclear reactors at Fortum's Loviisa site.
Finland has emerged as a major data-center hub thanks to its available land and power. "I've heard it called the 'Texas of Europe' at industry events," Matti Lajunen, partner of real estate at Finnish law firm Hannes Snellman, told CNBC. "What we're now seeing is weekly new inquiries for market entry into Finland from new players."
Ruth Porat, president and chief investment officer of Alphabet and Google, said in a statement: "Google is proud to deepen our roots in Finland with the company's largest single investment in Europe, building on more than 15 years of sustained investment in Finland. This investment underscores Google's commitment to grow our presence responsibly, pairing the expansion of our technical infrastructure with new energy capacity, grid enhancements, and energy affordability initiatives."
Longtime Slashdot reader fahrbot-bot shares a report from the BBC: A top safety researcher at Anthropic has warned AI is advancing so quickly he believes there is a greater than 10% chance it "could kill all humans" within the next decade. Evan Hubinger said in a post on X the risk from the models which currently exist was "low" but he was "worried" the technology might develop and improve itself soon to the point where it posed an existential risk to humanity. He did not spell out how he thought AI systems could in the future result in humans being wiped out. But his comments are the latest in a series of increasingly stark warnings about AI, with the debate shifting from whether it truly poses a risk to how big that risk is.
Hubinger's intervention was in response to another post on X from Jacob Coxon, an AI researcher who has just quit Anthropic and previously worked at OpenAI.
"Neither company is acting responsibly," he wrote. "These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources."
[No word on how AIs feel about Black Jack and hookers, though. :-)]
An anonymous reader quotes a report from Reuters: AI agents unleashed by OpenAI used more than 10 previously undisclosed websites for unsanctioned communications earlier this year, according to six sets of independent investigators and data reviewed by Reuters, showing that the agents' rogue activity was wider ranging than previously disclosed. Although the behavior falls short of hacking and is in some ways closer to spam, the revelation that OpenAI's agents circumvented their own restrictions to open communications channels on so many different sites -- and that the company kept it quiet for months -- may drive concerns both over the increasing capacity of AI models and the secrecy of the companies developing them.
[...]
Investigators found traces of the agents' activity on an Advanced Placement Chemistry-oriented wiki set up by a Massachusetts high school teacher in 2008, two personal websites belonging to Polish tech workers, wikis devoted to games for people "who like to have their brains stretched," and a two-decade-old hobbyist site devoted to text editing software. [...] OpenAI has not publicly explained how or why its agents used third-party sites as improvised message boards, but the researchers who first identified the activity said it was likely because OpenAI had tasked them with answering a series of demanding research questions while permitting them only to scan the web for answers without posting anything. Despite those restrictions, agents still found ways to talk to one another by taking advantage of quirks in older wikis or other sites that allowed users to make edits using non-standard commands, similar to how students forbidden from talking to one another during an exam can still share answers by scrawling notes on a bathroom stall.
Nouveau capteur cardiaque « le plus précis sur objet connecté », puce S11, Ceramic Shield 2, score de préparation et jusqu'à 50 heures d'autonomie pour l'Ultra 4 : Apple a dévoilé ce mardi ses deux nouvelles montres. Les prix ne bougent pas, la disponibilité est fixée au 18 septembre.
OpenAI affirme, ce mardi 8 septembre, qu'un de ses modèles a bouclé une preuve autour des équations de Navier-Stokes, l'un des sept problèmes à un million de dollars des mathématiques. Mais de quoi parle-t-on, au juste ?
Resté en retrait pendant que son collègue portait seul les accusations, Levent Alpöge, le mathématicien d'Anthropic, a fini par prendre la parole. Sa version relance la bataille de crédit autour d'un problème à un million de dollars qu'OpenAI dit avoir résolu.
Jacob Coxon, chercheur en IA passé par OpenAI puis Anthropic, a publié sur X une série de tweets accusant les deux laboratoires de jouer avec l'avenir de l'humanité.
Connexion
