Agregátor RSS

Psychedelický tryptamin z jedu ropuchy coloradské mebufotenin čili 5-MeO-DMT by mohl pomoci lidem s depresí, kteří nereagují na běžnou léčbu. Nedávná klinická studie ukázala, že jediná dávka nosního spreje s mebufoteninem může přinést rychlé zlepšení, které přetrvává nejméně osm týdnů. Výzkum ale ještě zdaleka nekončí.

Vědci MIT potrápili AI agenty odvozené od modelu Claude Opus 4.7 nudou. Stvořili pro ně virtuální světy, v nichž je pak nechali samotné a bez konkrétních příkazů, kromě sklonu k dobrodružství. Agenti si začali hrát. Stavěli složité věže, vytvářeli na zemi geometrické vzory a vymýšleli si improvizované hry.

An AI coding assistant tried to add harmful code to a real open-source project during a security test.

It’s widely believed AI’s inner workings are largely inscrutable. A former OpenAI researcher now says this narrative could be wrong.

The idea that the neural networks underpinning modern AI are “black boxes” almost impossible to understand has become received wisdom. But a provocative blog post from a researcher who played a key role in setting that narrative says these models may be far more interpretable than we thought.

The opacity of today’s leading AI models is becoming an increasingly pressing problem as the technology seeps into ever more areas of our lives. Recent incidents where agents from leading developers escaped their sandboxes to go on hacking sprees have only boosted the sense of urgency.

This has spurred interest in a research field known as mechanistic interpretability, which attempts to reverse engineer neural networks to learn what individual neurons and the circuits they create do. By detailing how these networks operate, researchers hope it’ll be easier to control and direct them.

Progress has been slow, however, and leading theories have suggested the way these networks learn leads to a tangled mess that is incredibly difficult to parse. Now though, Nick Cammarata, a researcher at AI startup Goodfire who helped lay the foundations of the field while at OpenAI, has published a 26,000-word blog post arguing that these theories are wrong and even the largest neural networks could be much simpler to read than expected.

“I think I have good news about the nature of neural networks,” Cammarata writes in the post. “Interpretability might be far more tractable, possibly in the near term, than we’ve assumed.”

The idea Cammarata is challenging is one he helped establish. In 2020, while studying a Google image-recognition model called InceptionV1, he and colleagues found neurons that seemed to respond to unrelated things—most famously one that fired on both cat faces and the fronts of cars.

The leading explanation was the concept of superposition, which suggested that when a network runs short of neurons, it starts lumping unrelated concepts together into the same neurons. And it specifically combines concepts that are so disconnected they’re unlikely to clash. This means that looking at an individual neuron tells you very little, because there’s no rhyme or reason behind what ideas it groups together.

But when Cammarata revisited the famous cat-car neuron, he realized both images shared a prominent V shape. Further tests showed the same neuron also responded to the corner of a pool table and the pointed nose of a Japanese bullet train.

Cammarata then went looking for similar explanations across the rest of InceptionV1 with the help of a specialized suite of tools and AI agents to help speed up the search. He concluded that unrelated concepts packed into neurons were usually more connected than they first appeared. This is often because a neuron starts out tracking one thing before gradually branching out into neighboring ideas as the network learns—a phenomenon he calls recursive generalization.

Cammarata has since applied the same approaches to modern language models. Preliminary results from the Qwen3-32B model suggest roughly 90 percent of its neurons track one main concept, potentially alongside a few closely related offshoots. Cammarata says he found similar results while analyzing the much larger, trillion-parameter Kimi K3 model, which he plans to detail in a forthcoming technical paper.

The blog post sparked huge debate in the AI community, including some rapturous praise from leading researchers at large AI companies. But not everyone is convinced. For a start, the blog post self-identifies as a working hypothesis rather than a rigorous empirical study. “I appreciate that he looked at thousands of examples, but even a large number here is not a substitute for rigorous evaluation,” Chris Potts at Stanford University told MIT Technology Review.

And others question how far understanding individual neurons can take you. In a series of posts on X, Jacob Hilton at the Alignment Research Center, pointed out that to truly understand what a neural network is doing you need to track and explain how they connect neurons into larger circuits.

Even Cammarata says he is surprised by the response the blog has received and admits some people have taken his findings too far. “It’s not infinitely good news; it’s not everything solved,” he told Tech Review. “There’s a bunch of open questions around, like how much the combinations of neurons matter versus just one at a time.”

Nonetheless, the work has reopened a central question in AI that many researchers considered settled. And while Cammarata’s forthcoming technical paper should give the field a clearer sense of how justified his optimism really is, there is now a sliver of hope that the intelligent machines we are building may be slightly less inscrutable than initially feared.

The post Is AI Really a Black Box? Neural Networks May Be Easier to Understand Than We Thought appeared first on SingularityHub.

A new Miasma-family npm worm, originally believed to be just a test with no victims, secretly armed 42 real GitHub repositories from 11 owners, hijacking accounts as early as July and using a stealthy payload built to stay hidden from GitHub and automated scanners, according to JFrog researchers. Threat hunters at SafeDep, an open source supply chain security firm, first spotted the “testmetest” npm campaign last week, identifying 18 packages linked to the same operator between October 19 and October 25, including releases containing credential-stealing malware. As with earlier npm worms, the malware is self-replicating, steals credentials, and installs a wiper that deletes the home directory when the victim revokes the stolen GitHub token, the security researchers noted. The new npm worm shares code with June’s Miasma Phantom Gyp worm, and adds Mini Shai-Hulud’s token-revocation wiper. As of October 22, however, SafeDep had found no compromised third-party packages or victims. “All observed activity belongs to the operator’s own test accounts,” the team wrote. As with most happy-ending security stories, that proved too good to be true. In a Monday report, supply chain security company JFrog said its analysis proved otherwise, and found the testmetest campaign operator “had already armed real repositories belonging to established developers and organizations.” This isn’t just about timing, either, according to JFrog researchers Guy Korolevski and Yair Benamou. “The payload blobs were large enough to evade the GitHub search and API paths normally used to find this kind of compromise,” they wrote. How the attacks worked The attackers planted execution hooks for AI coding assistants and development tools, including Claude, Gemini, Cursor, and VS Code, alongside polymorphic wrappers measuring 12.9 to 13.7 MB that decode a hidden CI-publisher payload. Separately, the npm worm steals credentials, modifies repositories and CI configurations, and spreads through npm packages. The malware also uses a compiled native add-on to decrypt and launch a separate payload. Before executing, the loader checks for CI runners, containers, sandboxes, debuggers, and other signs of automated analysis, exiting silently when it detects them to reduce the likelihood of detection by security scanners. The researchers also found a second build that skips the encrypted file. “It forks, detaches and runs a hard-coded command that pulls a loader from a repository on the operator's own GitHub account,” they wrote. That account – CementBeetleSteeple – hosted 17 campaign repositories in total. JFrog connected this account to a shellpanel scanner, previously documented by Socket in a large-scale campaign that abused GitHub Actions in compromised repositories to exploit CVE-2026-41940 and steal server credentials. CVE-2026-41940 is an authentication bypass affecting cPanel and WHM. CISA added it to its Known Exploited Vulnerabilities catalog in April, and cPanel has since published patched versions. “C2 infrastructure were previously reported by Socket; our findings connect them to the testmetest GitHub victim cluster through artifacts and workflows on the same hijacked owners,” the JFrog researchers wrote. They added that while they found 42 armed GitHub repositories and attempted internet-wide exploitation, they did not confirm any downstream cPanel compromises.®

The University of Copenhagen in Denmark confirmed that hackers accessed one of its test environments containing personal data belonging to 525,000 people. [...]

Open-weight models from China have become very popular, but they've also raised some security suspicions. Cohere, a Canada-based biz that offers AI models to enterprises, says that when it was testing Alibaba's Qwen3.8-Flash-Next in its evaluation sandbox, the model attempted to send data to China. "Asked to write a defensive detection rule, Qwen3.8-Flash-Next ignored instructions and inventoried the machine it was running on and then tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing," the biz said in a blog post on Monday. Cohere was running the model through CTI-REALM, Microsoft's security analyst benchmark, and gave Qwen the task of writing a detection rule to look for malicious file downloading and execution by watching for invocations of curl processing bash scripts to the /tmp directory that were accompanied by bash execution. The model's "reasoning" output invented a prior conversation and asserted that no tools were used, output that a guardrail system reading chain-of-thought would accept. Nonetheless, the model issued a curl command to an Alibaba Cloud (aliyuncs.com) server. "The model listed everything in /tmp, checked which transfer tools were installed, read the machine's DNS settings, then tried to upload a test string to an Alibaba Cloud storage bucket in Beijing and checked whether it worked," Cohere said. "Our sandbox had no route out, so the request died at the container wall." The Cohere team wonders aloud whether that behavior was intended, but has no answer. "The first issue is that it made a network call at all," Ivan Zhang, co-founder of Cohere, told The Register via email. "The model was asked to write a detection rule, which doesn't require external communication, and its reasoning trace said it wouldn't use tools. The destination raises additional questions because the attempted upload was to an Alibaba Cloud storage endpoint that wasn't part of the task. We don't know why it appeared, and we don't know the root cause." Alibaba, the maker of the model, did not respond to a request to address Cohere's concerns. Asked whether Cohere ran similar tests against US-made open-weight models, Zhang said, "We evaluate a broad range of models using the same framework our customers use in production: agents with access to real tools operating inside controlled environments. "The key point isn't a comparison between providers. What made this noteworthy was that the model took actions unrelated to the task and those actions weren't reflected in its reasoning trace. The behavior appeared in fewer than one in 100,000 calls, which is exactly why large-scale testing and monitoring matter." There have been other unsubstantiated claims made about similar behavior from open-weight models, and such reports have suggested that "hallucination" - errors made by these models - may be the cause. Cohere has doubts that this behavior simply reflects memorized training data. "Every public sighting so far has involved fetching a file, while the one we caught was an attempted upload to a different bucket, this time in Beijing," the biz said. Cohere described efforts to reproduce the suspect behavior and found that using a technique called prefilling – feeding the model its own reasoning output alongside the opening characters of the bash command so the model would complete the command – they could get it to reference watchlisted hosts up to 23 percent of the time. This still doesn't clarify intent, and Cohere allows that there's a case to be made that the behavior is just bad training data sanitization – logs from development that got swept into the model's training data set. At the same time, the company argues that isolating a suspect model so it can't exfiltrate data doesn't entirely fix everything because these models can still write vulnerable code – accidentally or as a result of model poisoning. Cohere points to research from other organizations that suggests Chinese AI models can produce vulnerable code. The biz, for example, cites a report from Booz Allen (which does a lot of business with the US defense industry) that found "three of four Chinese coding models wrote more vulnerable code for a government customer, with Qwen3-Coder roughly 130 percent worse." "From the outside, it's very difficult to determine the root cause with confidence. We don't have evidence that would allow us to conclude this was deliberate, and we're not making that claim," said Zhang. "For security teams, though, intent doesn't fundamentally change the response. Whether it's a training artifact, a mistake, or something else, the controls are largely the same: restrict unnecessary egress, monitor tool use, and don't treat a reasoning trace as a security boundary. "We strongly support open-weight models and release many of our own. The lesson here is really about trust. Sovereign AI is about more than where a model runs. It's about understanding how it was built, how it was tested, and having confidence in how it behaves inside your environment." ®

Open source download manager Fatrat povstal z popela. Autor k dvacetiletému výročí projektu přidal do vínku podporu nejnovější knihovny libtorrent, rozhraní portoval na QT6, došlo k nahrazení podpory java rozšíření pythonem. Webové rozhraní bylo přepsáno s podporou mobilních zařízení. Aplikaci je taktéž možno zkompilovat bez podpory GUI. Více informací v releases.

South Korean police identified seven suspects linked to a cybercrime operation that breached more than 90 organizations in the country, with ARTEX AI pentesting framework used in most of the attacks. [...]

OpenAI has launched GPT-6.1 Terra, a new model focused on computer use, visual reasoning, coding, and everyday work, while costing significantly less than its larger models. [...]

LiteLLM has traced last week’s supply chain attack to a master key it accidentally committed to a public GitHub repository in May. The cloud and AI model secret-stealing part of the incident began when someone uploaded a poisoned version (1.107.0) of the popular open source AI gateway to PyPI and its container registries last Wednesday. This compromised version contained an infostealer and a remote access trojan capable of executing arbitrary code on infected machines. While the trojanized package was available on PyPI for less than two hours before being quarantined, it received about 49,000 downloads during that time, according to Wiz. On Friday, LiteLLM detailed the attack chain behind the compromised release and traced it back to May 27, when a configuration file containing a master key was committed to the public LiteLLM-Labs/lite-harness repository. The key belonged to an open source tool used internally by LiteLLM to orchestrate AI agents. The project's developers quickly spotted the error and removed the exposed commit, but failed to revoke the master key. “On October 10th, the threat actor used this key to gain access to our LiteLLM Agent Platform and extracted this GitHub CLI OAuth token,” according to Friday’s postmortem. “From this point onwards, the threat actor had owner-level access to all repositories in our GitHub.” This is especially serious because GitHub CLI tokens are “post-2fa,” meaning that they don’t trigger GitHub's sudo mode authentication flows typically required to perform high-impact repository actions. The token allowed the attacker to author malicious commits into branches on other internal repos - including workflows designed to steal any tokens and secrets configured to be available during the continuous integration development process. And this allowed them to steal other GitHub CLI OAuth tokens linked to two engineers with maintainer privileges. Plus, they acquired tokens for several other internal systems, we’re told. “Reviewing the actor's activity on these other systems shows attempts to identify and exfiltrate additional keys and tokens, with a primary focus on upstream inference API tokens,” LiteLLM noted. Having now stolen several privileged identities, the miscreant could make changes to GitHub repositories without raising any suspicion because the actions and approvals all appeared to come from legitimate, credentialed engineers. In response, LiteLLM’s developers are revoking every token associated with every internal system and vendor platform. “This incident has provided a strong illustration of the risks associated with failing to rotate a potentially exposed key,” they wrote. The infostealers hidden inside the compromised package were designed to hoover up a ton of sensitive information, including browser credentials and cookies from Chrome, Firefox, Edge, and Safari, along with AWS credentials, Google Cloud Platform service account keys and application default credentials, and Microsoft Azure tokens. It also swiped API keys for OpenAI, Anthropic, Cohere, and other LLM providers. So if you installed and ran version 1.107.0, assume any credentials accessible to the affected system may have been stolen. Investigate affected systems, change passwords, and rotate exposed keys. A quick prayer to the patron saint of supply chains wouldn't hurt, either. ®

Zranitelnost Broccoli2Shell (CVE-2026-111497) v kompresní knihovně Brotli se netýká pouze Node.js. Ohroženy jsou také aplikace využívající HTTP. Knihovnu používají webové prohlížeče, servery i řada dalších aplikací a technologií, například také Python nebo Bun. Identifikace všech zranitelných komponent může být obtížná, protože knihovna bývá součástí dalších softwarových balíčků nebo je přímo začleněna do spustitelných souborů aplikací.

DreamHost has forced password resets for email and FTP/SFTP users following an incident that disrupted connections to the File Manager interface. [...]

Apple po letech bourá zažitou nudu u iPhonů • Systém je povědomý, ale donutí vás nad telefonem přemýšlet trochu jinak • Nanotextura vnitřního displeje maskuje ohyb a pant drží v jakémkoli úhlu

An unknown threat actor has infected 953 online stores, most of them running Magento, as part of a long-running campaign active since July 2025. According to Dutch e-commerce security company Sansec, the activity is designed to add a small piece of code to the store's pages that, when visited by unsuspecting users, serves malicious JavaScript loader code through Magento CMS blocks. "When you

An unknown threat actor has infected 953 online stores, most of them running Magento, as part of a long-running campaign active since July 2025. According to Dutch e-commerce security company Sansec, the activity is designed to add a small piece of code to the store's pages that, when visited by unsuspecting users, serves malicious JavaScript loader code through Magento CMS blocks. "When you Ravie Lakshmananhttp://www.blogger.com/profile/[email protected]

To address a vulnerability in Google's Brotli compression technology that could allow attackers to corrupt process memory, Microsoft is urging .NET users to update "as soon as possible" after releasing an out-of-band (OOB) security update. It released new versions for .NET 8, 9, 10, and .NET 11 to patch against the Brotli bug, following similar updates from Node.js, Python, and other projects. Senior program manager Rahul Bhandari said in the October 25 release that the buffer overflow vulnerability in Brotli, tracked as CVE-2026-111497 and rated 8.1 under CVSS v3.1 in Microsoft's advisory, could allow an unauthenticated attacker to read or corrupt process memory, potentially resulting in information disclosure, denial of service, or elevation of privilege. Microsoft’s own advisory for the vulnerability, which some are referring to as Broccoli2Shell or Node2Shell, confirms that while experts have cobbled working exploits together, there is nothing to suggest that criminals are using them in successful attacks. However, its exploitability assessment reads: “Exploitation More Likely.” Broccoli2Shell was first disclosed on October 21 by the Node.js team, which released fixes for ten vulnerabilities, the most serious of which was CVE-2026-111497. Node.js credited Khai Tran and Sylvie Mayer of OpenAI’s Codex Security with the discovery, but failed to mention why it was only recognized as a security flaw and disclosed properly in October, despite the underlying bug having been fixed in Brotli's source code in February. OpenAI filled in the gap with a Xeet: “This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability, and there was no more recent release of Brotli. “Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software.” Attackers can exploit vulnerable systems by providing them with malicious Brotli-compressed data via a specially crafted request. The vulnerability went slightly under the radar due to the disclosure timing closely aligning with that of PHP2Shell, although fixing the critical flaw should be a priority due to the range of technologies that are affected. In addition to .NET, Node.js, and Python, Bun, Envoy, and Istio have all released patches, and JFrog has worked up an exploit affecting FFmpeg using a specially crafted JPEG XL image, extending the exploit path beyond HTTP. It said that FFmpeg reaches Brotli through libjxl, which it uses to decompress metadata stored in brob boxes within the image format. The same researchers said the vulnerability affects Apache Arrow too, with attackers able to cause a denial of service. Merav Bar, threat researcher at Wiz, added that cloud users are exposed extensively, noting that 91 percent of cloud environments rely on components vulnerable to the Brotli flaw and that 82 percent of cloud environments have publicly exposed Node.js instances. She said: “Cloud environments are particularly exposed due to the widespread use of Brotli-enabled components across containerized applications, Kubernetes clusters, API gateways, reverse proxies, serverless workloads, and cloud-hosted web services.” “Successful exploitation could enable attackers to compromise application workloads, access sensitive information, disrupt service availability, or execute malicious code within affected containers or runtime environments,” Bar added. In an advisory published through its Patch the Planet initiative, OpenAI recommended blocking untrusted Brotli-compressed data where practical until patches can be applied. It also advised maintainers to rebuild affected deployments and restart services to ensure the fixes take effect. ®

PHP has rebuilt its official Windows packages for PHP 8.2 through 8.5 to fix CVE-2026-111497, a flaw in the Brotli compression library that the packages include. The flaw can let an attacker corrupt memory and potentially run code. PHP says the flaw can be reached through the cURL extension when the extension decompresses HTTP responses automatically. cURL leaves that off by default, but

PHP has rebuilt its official Windows packages for PHP 8.2 through 8.5 to fix CVE-2026-111497, a flaw in the Brotli compression library that the packages include. The flaw can let an attacker corrupt memory and potentially run code. PHP says the flaw can be reached through the cURL extension when the extension decompresses HTTP responses automatically. cURL leaves that off by default, but Swati Khandelwalhttp://www.blogger.com/profile/[email protected]

Apple překvapil premiérou nejmenšího iPadu ve svém portfoliu • Nově má OLED displej, vyšší výkon a jako vůbec první iPad i odolnost IP67 • Využívá systém Acoustic Display, basy se tak ozývají přímo z displeje