Connexion
OpenAI will begin automatically watermarking text generated with ChatGPT in the European Union, the company has announced. It will also offer the watermarking feature in other regions, but it will be off by default outside of the EU.
The move in Europe is driven by a need to comply with the EU AI Act, which took effect in August. It requires marking content produced by AI models in a way that another tool can detect. Unfortunately, there is still no completely effective and reliable way to do that. A few standards already exist, like SynthID and the C2PA project, but they are relatively easy to circumvent for anyone with basic know-how.
The same is likely true for OpenAI's watermark. Its method is proprietary; the company calls it textGrain, and has published a technical paper explaining how it works. But in general, it works like other LLM watermarking tools we've seen in the past: It puts patterns in the word choices that are not clear to a human reader and that don't meaningfully change the general quality of the output, but that someone with a key can use a specialized detector to find. OpenAI says it will be giving access to the detector to a limited number of researchers and organizations and providing a request-for-approval process for others to be added over time.
Hackers have been exploiting a critical vulnerability in the Zimbra Collaboration Suite in an attempt to obtain email backups and authentication credentials of vulnerable organzations, Microsoft has warned.
The vulnerability, tracked as CVE-2026-73570, lets attackers remotely issue operating system commands without authentication. Zimbra maintainer Synacor issued a patch on July 20, but didn’t disclose the vulnerability for more than three weeks after that. The security-focused Shadowserver Foundation said last week that its scans found that 274 separate instances of the Zimbra Collaboration Suite had been compromised. The number of servers running the software has fluctuated from 19,000 in the week following the patch to about 12,000 in the weeks following that. Currently, Shadowserver is tracking about 10,000 instances.
From July 28 to August 7, Microsoft said Wednesday, the company detected two distinct scanning tools probing the Internet for vulnerable endpoints. The attackers first validated their exploit worked by sending HTTP, requests and DNS, ICMP, and out-of-band identity checks to domains hosted on public services. The probes allowed the attackers to confirm the exploit successfully executed commands on vulnerable servers without actually compromising them. Eventually, the attackers began using their command injection capability to install malicious payloads. Microsoft wrote:
The genetic code is what life everywhere uses to convert the information contained in DNA into specific protein sequences. With minor variations, the same genetic code is used by every living thing on Earth, suggesting it was already present in the last common ancestor of all of them. It's not an easy thing to change, because so many things in every cell depend on it.
Nevertheless, some preliminary steps have been taken. Researchers have managed to add some new amino acids to a bacterial cell and were able to make proteins that were one amino acid less than usual. But it's a slog; for some of this work, people have had to re-engineer every single gene in a bacterial genome.
Now, researchers have found a way to operate two separate genetic codes simultaneously, avoiding the need to do any work to compensate for altering the code that every protein in a cell relies on. They didn't test it in an actual cell, and it might cause some problems there. But it's a creative solution that should accelerate some synthetic biology work.
Cloudflare has open-sourced its Cloudflare OS platform, which it first developed as an internal workspace for employees to build apps using AI agents—including people who are not software developers or engineers. The company also touts a security framework designed to reduce the risk of employee vibe-coding sessions creating serious security flaws or leading to data breaches.
The tech company spent several months building and internally testing Cloudflare OS, which allows employees to describe workflows in natural language so that an AI agent can code them into applications. In an August 5 blog post announcing the open source version’s availability on GitHub, the company claims thousands of Cloudflare employees use the platform on a daily basis to “create documents and slides, automate repeatable tasks, and build small apps to visualize data and help them do their work.”
“This is a full-on personal app vibe coding platform, in which the sandbox is so secure that you can pretty much go wild—the AI cannot introduce a significant security bug,” said Kenton Varda, principal engineer at Cloudflare, in a post on the social media platform X. “We believe a company's security team can feel comfortable giving non-technical users permission to vibe code and then sleep soundly at night.”
A recent examination of hundreds of mobile apps marketed toward US military personnel found more than one in eight contained software built by companies in China, Russia, or other foreign nations, raising fresh concerns that adversary governments could harvest data revealing where service members live, work, and deploy.
According to researchers at Purdue University, the US Military Academy at West Point, and Florida International University, one popular app used by service members to rate living conditions on their own bases include code from Huawei, the Chinese telecom that US regulators flagged as a national security threat in 2020. Two others were built by Russian companies and incorporate the Russian ad service Yandex.
The largely unregulated advertising industry that tracks Americans online treats civilians and service members mostly the same—unless there is profit in telling them apart—despite evidence that exposure can reveal troop deployments, unit movements, and the routines of personnel within intelligence facilities and hardened shelters where nuclear weapons are believed to be stored.
There are a lot of AI coding applications out there, and as impressive as large language models and the agents they enable have become, many of the most recent developments in AI-assisted development have been in the software that manages those models, not just the models themselves.
Earlier this summer, I spoke with the head of product for Claude Code, Anthropic's Cat Wu, about that company's approach to building that software.
Last year, when we tested out the "Agent Mode" in OpenAI's Atlas web browser, we complained that any automated tasks tended to stop after a few minutes, limiting its usefulness for ongoing or complex tasks. With today's release of ChatGPT Work, OpenAI says it has solved that problem with a new tool that can "stay with a project for hours if needed, and turn a goal into finished work."
The company is challenging users to evaluate ChatGPT Work by "giv[ing] it a task you already know well," such as analyzing a budget or preparing a sales meeting. The company also promises that ChatGPT Work can automate entire workflows, going from customer research to a campaign brief to locally tailored marketing assets, for instance. At the same time, the company stresses that the tool will wait for you to "approve important actions."
ChatGPT Work also integrates Scheduled Tasks, a souped-up version of cron jobs that can "take repetitive tasks off your plate" on a schedule or whenever a monitored event occurs. These tasks can keep going when you're away from your desk and can be monitored from your phone, the company says.
Anthropic quickly removed a tracker secretly monitoring Claude Code users in China after a security researcher exposed the hidden code and condemned the spyware-like tracking as a “serious breach of user trust.”
Last week, a web developer known as “Thereallo” was researching privacy issues in Claude Code and was shocked to find that the AI firm was using “prompt steganography” to hide code that tracks Chinese users “in plain sight.” This code wasn’t malicious, but it was sending information to Anthropic that most users wouldn’t detect, relying on shorthand markers to quietly flag users’ timezone, proxy, and potential connection to Chinese AI labs that Anthropic has accused of distillation attacks.
On X, Anthropic engineer Thariq Shihipar confirmed that the tracker was added to Claude Code as an “experiment” in March. According to Shihipar, the code “was meant to prevent account abuse from unauthorized resellers and protect against distillation.” Regarding the former, The Washington Post found unauthorized retailers have sold access to free models for $1 a month, and pro subscriptions that can cost $100 monthly sell for "as little as $12."
OpenAI, the company behind ChatGPT and Codex and the models those tools use, and Broadcom, an established silicon supplier, have announced a new chip, called Jalapeño, designed specifically for large language model inference in data centers.
The companies intend to deploy the chip at large data centers and claim this is just the first generation in a long-term project that will see chips refined over time.
Operating system makers take many steps to prevent their wares from accepting commands from remote devices. The safeguards, designed to thwart malicious attacks, typically require hackers to jump through all kinds of hoops to bypass the measures. But what if remote code execution were as simple as being within Bluetooth range of a speaker connected to the targeted device?
It turns out it can, at least when the speaker is a Sound Blaster Katana V2X sold by Singapore-based Creative Technologies. The speaker, which sells for $283, is widely acclaimed with numerous reviews showering praise on the sound and performance of it and its predecessor, the Sound Blaster V2.
Researcher Rasmus Moorats stumbled on the hack by accident, after he purchased a Katana V2X, a soundbar that connects to PCs, Macs, and Linux devices over USB or Bluetooth. Moorats was curious if he could create a Linux tool that communicated with his speaker. He discovered he could do so through CTP, a proprietary mechanism he guesses is short for Creative Transport Protocol.
SAN FRANCISCO—Amid an ever-expanding array of surfaces, growing demand for tokens and compute, and a rapidly evolving user base, Anthropic doesn't have a long-term road map for Claude Code. However, it's betting that such a plan would be rendered moot by improvements in model capabilities and new signals from developers on how best to use it. That's the takeaway from a 30-minute conversation Ars had with Cat Wu, Anthropic's head of product for Claude Code.
Last week, in a three-level car rental parking garage meticulously converted into an event space in downtown San Francisco, Anthropic put on its second annual Code with Claude developer conference. As previously reported, the single-day event included a keynote introducing new features for Managed Agents and announcing a compute deal with SpaceX.
That compute deal was accompanied by a doubling of usage limits for Claude Code users on the company's Pro and Max plans—a response to a lot of user frustration about a compute crunch, especially in recent weeks.
SAN FRANCISCO—At its Code with Claude developer conference on Wednesday, Anthropic announced a deal with SpaceX to utilize the entire compute capacity of the latter's data center in Memphis, Tennessee.
On stage at the conference, CEO Dario Amodei said the deal was intended to increase usage limits for Anthropic's Pro and Max plan subscribers.
The announcement was accompanied by an increase in those usage limits; Anthropic doubled Claude Code's five-hour window limits for Pro and Max subscribers, removed the peak-hours limit reduction on Claude Code for those same accounts, and raised API limits for its Opus model. The table below outlining the Opus changes was shared in the company's blog post on the topic.
SAN FRANCISCO—At its Code with Claude developers' conference, Anthropic has introduced what it calls "dreaming" to Claude Managed Agents. Dreaming, in this case, is a process of going over recent events and identifying specific things that are worth storing in "memory" to inform future tasks and interactions.
Dreaming is a feature that is currently in research preview and limited to Managed Agents on the Claude Platform. Managed Agents are a higher-level alternative to building directly on the Messages API that Anthropic describes as a "pre-built, configurable agent harness that runs in managed infrastructure." It's intended for situations where you want multiple agents working on a task or project to some end point over several minutes or hours.
Anthropic describes dreaming as a scheduled process, in which sessions and memory stores are reviewed, and specific memories are curated. This is important because context windows are limited for LLMs, and important information can be lost over lengthy projects. On the chat side of things, many models use a process called compaction, whereby lengthy conversations are periodically analyzed, and the models attempt to remove irrelevant information from the context window while keeping what's actually important for the ongoing conversation, project, or task.
The genetic code is central to life. With minor variations, everything uses the same sets of three DNA bases to encode the same 20 amino acids. We have discovered no major exceptions to this, leading researchers to conclude that this code probably dated back to the last common ancestor of all life on Earth. But there has been a lot of informed speculation about how that genetic code initially evolved.
Most hypotheses suggest that earlier forms of life had partial genetic codes and used fewer than 20 amino acids. To test these hypotheses, a team from Columbia and Harvard decided to see if they could get rid of one of the 20 currently in use. And, as a first attempt, they engineered a portion of the ribosome that worked without using an otherwise essential amino acid: isoleucine.
First off, why would you do this? Most work in the field has focused on altering the genetic code in ways that are useful, such as using more than 20 amino acids to enable interesting chemistry.
The system prompt for OpenAI's Codex CLI contains a perplexing and repeated warning for the most recent GPT model to "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query."
The explicit operational warning was made public last week as part of the latest open source code for Codex CLI that OpenAI posted on GitHub. The prohibition is repeated twice in a 3,500-plus word set of "base instructions" for the recently released GPT-5.5, alongside more anodyne reminders not to "use emojis or em dashes unless explicitly instructed" and to "never use destructive commands like 'git reset --hard' or 'git checkout --' unless the user has clearly asked for that operation."
Separate system prompt instructions for earlier models contained in the same JSON file do not contain the specific prohibition against mentioning goblins and other creatures, suggesting OpenAI is fighting a new problem that has popped up in its latest model release. Anecdotal evidence on social media shows some users complaining about GPT's penchant for focusing on goblins in completely unrelated conversations in recent days.
Lateo.net - Flux RSS en pagaille (pour en ajouter : @ moi) 