AI AND ML

Researchers urge developers to see that less is more when it comes to instructions

If you’re exposing your agent to a strong odor, it’s time to clean up your instructions.

Risky or poorly structured code patterns are known as “code smells,” and it turns out coding agent directives can be similarly redolent, leading to wasted tokens and worse output. 

Coding agents rely on configuration files that summarize expected agent behavior. These context-enhancing files are commonly written in Markdown and named either CLAUDE.md for those using Anthropic models or AGENTS.md for pretty much everyone else.

They include various text instructions that advise the coding agent about desired behavior and tool use. And they can get rather wordy. Anthropic advises no more than 200 lines of text because longer files consume model context and may hinder model coherence.

Researchers affiliated with the computer science department of the Federal Institute of Minas Gerais in Brazil recently scoured some 532,000 files to build and analyze a dataset of 100 popular open-source projects containing either an AGENTS.md or a CLAUDE.md file.

“Our results show that configuration smells are widespread,” the authors state. “Lint Leakage was the most common smell, affecting 62 percent of the files, followed by Context Bloat (42 percent) and Skill Leakage (35 percent).”

Linting is the process of running automated tools to check code for programming and style errors. Lint Leakage refers to agent instructions that repeat rules already enforced by linters, format checkers, and static analysis tools. Duplicative rules waste tokens by burdening the underlying model with guidance for a task already handled reliably by programmatic tools.

Context Bloat, as its name suggests, describes the tendency of developers to overspecify code agent behavior. “Bloated configuration files increase token consumption, raise costs, and reduce the visibility of important instructions,” the authors observe, pointing to Anthropic’s recommendation of no more than 200 lines of text.

Skill Leakage, another common configuration smell, occurs when rarely used tools or practices get added to the AGENTS.md file, which gets loaded in every agent session. The agent instructions would be better in a separate skills file (e.g. SKILLs.md) that gets loaded only when needed. Skill leakage also expands the agent’s context unnecessarily and potentially distracts agents from other things.

Other agentic odors include: Blind References, which happens when configuration files reference external documents (e.g. via URLs) without explaining when that resource becomes relevant; Init Fossilization, configuration details set up upon a project’s initialization that are no longer relevant; and Conflicting Instructions, which occur when agent directives contradict each other.

The study authors say that they found at least one of these six smells in 91 of the 100 AGENTS.md files tested.

“These results suggest that developers could benefit from catalogs and tools designed to spot configuration issues in agent configuration files,” they conclude in the preprint paper, entitled “Configuration Smells in AGENTS.md Files: Common Mistakes in Configuring Coding Agents.” The authors are Helio Victor F. dos Santos, Vitor Costa, Joao Eduardo Montandon, Luciana Lourdes Silva, and Marco Tulio Valente.

The message here is that less is more when it comes to code agent configuration files, perhaps even to the point that anything is worse than nothing. 

Similarly, when ETH Zurich boffins examined the impact of context files for agents a few months ago, they found [PDF] that developer-generated instructions raised costs and only improved code performance about 4 percent, while LLM-generated instructions had a small (3 percent) negative impact on agent-generated code.

They concluded “unnecessary requirements from context files make tasks harder, and human-written context files should describe only minimal requirements.” ®

The ultralight may become a permanent fixture in Apple’s smartphone lineup.

Google has started rolling out Wear OS 7 to Pixel Watch users. This brings what is arguably the biggest software update of the year to the company’s smartwatch lineup.

The update introduces new Gemini-powered features, redesigned widgets, and battery life improvements. However, it won’t be coming to the original Pixel Watch.

According to Google, the rollout is now underway for the Pixel Watch 2, Pixel Watch 3, and Pixel Watch 4. Availability is expected to expand gradually over the coming days. Alongside a refreshed interface, Wear OS 7 is designed to improve efficiency. Google claims battery life could increase by up to 10%. This depends on how the watch is used.

One of the most noticeable changes is the shift from full-screen tiles to a new widget system. This system looks much closer to Android’s smartphone widgets. The update also adds live notifications. This allows users to see real-time updates directly on their watch. It works in a similar way to Android’s Live Updates feature.

Hudson Rock said the attackers went on to “actively intercept SSL VPN authentication hashes and crack them using a massive, dedicated 45-GPU cluster managed via Hashtopolis.” From there, they used the GPU cluster to crack the hashes, meaning to try massive combinations of plain-text passwords until they found the right one. These passwords allowed the threat actors to move laterally to compromise Active Directory environments and other centralized authentication systems.

“This aggressive methodology has led to severe, real-world consequences,” Hudson Rock said. “Diachenko’s research confirmed full network compromises at multiple organizations across Japan, Taiwan, Vietnam, Iraq, and Turkey. Most alarmingly, this includes a Turkish NATO defense contractor from which classified defense documents were successfully exfiltrated by the group.”

In the interview, Diachenko put it more succinctly. “The scale is the sophistication,” he said.

The scale didn’t stop there. The attackers used the massive cluster to run a” feedback-driven, 12-level recursive system.” In other words, there wasn’t a single flat dictionary run. Password candidates came from custom dictionaries with as many as eight words, common keyboard patterns, and cracking rules. Each one looped back with each step. When guesses were successful, the passwords were fed back as seeds to generate still more candidates. In other words, the cracking techniques improved with each successful guess.

“They were quite innovative on that,” the researcher said.

The innovation contrasts sharply with the operational security of the attackers, who left artifacts on the server they used. In hacker circles, such moves are considered amateur mistakes.

Hudson Rock said that the top countries where compromised devices were found were India, the US, Taiwan, Mexico, Turkey, and Thailand. The top industries affected were IT services, construction materials, telecommunications, construction and engineering, industrial equipment, and financial services. Other organizations whose data appeared in the database included: Foxconn, Samsung, Comcast, Siemens, PwC, and Accenture. Hudson Rock said that the database listed thousands of others, including major government agencies and critical infrastructure providers.

Firewalls have long been a favorite network entry point for hackers. These devices accept connections from the outside Internet, sit at the perimeter of a network, and have access to valuable resources deep inside.

The links above list a number of steps Fortinet firewall users should take to ensure their networks are secure. Given that the data has been available to cybercriminals and potentially other threat actors who, like Diachenko, found it, the risk is substantial.

Years after the events of Spider-Man: No Way Home, the latest trailer shows Tom Holland’s Peter Parker still living in the shadow of that memory-erasing spell. No one knows who he is anymore, not even his closest friends. The footage leans into that isolation while cranking up the personal stakes and physical chaos for the July 31 release.

The new teaser opens on a gritty New York City street, with Michael Mando’s Scorpion charging at Peter out of nowhere, dressed in a comic book-inspired outfit. The two engage in a massive, primal street battle. When Peter grips Scorpion’s tail, his eyes go completely black, a blink-and-you’ll-miss-it warning that he’s losing his sanity. Next thing you know, he’s flinging Scorpion into a car, gasping as if he’s losing his hold on reality.

LEGO Marvel Spider-Man vs. Mysterio: The Daily Bugle Building Toy – Pretend Play Set with 7 Minifigures…

• 4 ACTION-PACKED LEVELS – The Spider-Man vs. Mysterio: The Daily Bugle (76342) building toy lets kids ages 10 years old and up play out their Super…
• 7 LEGO MARVEL MINIFIGURES – Kids can stage epic battles with Spider-Man, Miles Morales, Ghost Spider, J. Jonah Jameson, Mysterio, Venomized Captain…
• PRETEND PLAY TOYS – This LEGO Marvel set includes a poseable Rhino mech with an opening cockpit, plus 3 transparent posing sticks and 2 transparent…

Then Peter’s mechanical web-shooters just fall apart at the wrists, and biological webbing shoots out in all directions. He’s swinging through traffic in a frenzy, catches a bird along the way, and then crashes into a sad couple getting married. It’s a cross between a wild, street-level disaster movie and a body horror film. Peter then runs to Bruce Banner (Mark Ruffalo) to figure out what’s going on with his DNA. Banner pulls out a device designed to keep the Hulk hidden and gives him a harsh warning: if Peter ever discovers him without that ‘thing,’ he should get out of there as soon as possible, since this hints at a whole bunch of gamma-powered problems waiting to burst.

The action kicks up again when Spider-Man clashes with the Hand, as the ninjas make a bigger mark in the MCU this time around. He’s spinning a giant web tornado across their ranks, and the skyscrapers behind him are crashing down. Then, just when you think it can’t get any wilder, Jon Bernthal’s Punisher appears in his combat van, seemingly stepping in to save MJ at one point, adding to the drama. The trailer then turns nuclear, with the most dramatic escalation yet occurring as a huge Grey Hulk appears, apparently under mental control. Peter merely stares up at the item and says, “Wait, what?” The Hulk became bigger?’ Then the two go toe-to-toe in a battle for the ages, destroying the city.

This second trailer takes the first and turns it up a level, really going deep into Peter’s mental struggle and the consequences of his lost identity. It still includes some street-level action, as well as larger MCU crossovers such as the Hulk and Punisher, but Sadie Sink’s role remains unknown for the time being. We have the famed Destin Daniel Cretton at the helm, working with Chris McKenna and Erik Sommers, who are back on writing responsibilities this time, and newcomer Justin Kuritzkes.

A newly discovered data leak dubbed “FortiBleed” has exposed what appears to be a collection of Fortinet and FortiGate VPN credentials for 73,932 firewall URLs at organizations worldwide.

The exposed data was first discovered by security researcher Bob Diachenko, who says he found a server containing what appeared to be valid Fortinet VPN credentials, including usernames, email addresses, and plaintext passwords.

According to screenshots and information shared by Diachenko, the database contains entries for Chevron, Samsung, Foxconn, Comcast, AT&T, Mercedes-Benz, Toyota, Sinopec, State Grid, and many others. 

“Massive Fortinet/FortiGate bruteforce/active exploitation campaign uncovered in action,” Diachenko posted on LinkedIn.

“Thousands of top vendors instances are listed in the files like this (see screenshot). This one alone has 21,634 domain names – from Chevron to Fortinet itself. All – with potentially working passwords to the FortiGate appliances obtained through various menas.”

The exposed data also included comments listing each organization’s industry, revenue, and number of employees, likely for planning attacks.

Diachenko later shared additional information that claimed the operation was conducted by a Russian-speaking multi-operator threat group that harvested credentials for FortiGate SSL VPN devices.

According to Diachenko’s investigation, the attackers allegedly conducted approximately 1.16 billion credential attempts against 320,777 FortiGate targets and an additional 2.1 billion attempts against 163,650 Microsoft SQL Server systems.

He further claimed the threat actors intercepted SSL VPN authentication hashes, cracked them using a 45-GPU cluster managed through Hashtopolis, and used the recovered credentials to move laterally into internal Active Directory environments.

Diachenko told BleepingComputer he obtained these details after analyzing additional files inadvertently exposed on the same server.

“They accidentally left an open directory with artefacts, connection strings, tooling, scripts and data online. Analytics obtained via their cron jobs, bash histories, logs etc,” Diachenko explained.

The researcher also stated that multiple organizations across Japan, Taiwan, Vietnam, Iraq, and Turkey were fully compromised, including a Turkish NATO defense contractor from which classified documents were allegedly stolen. 

Threat intelligence company Hudson Rock has since published its own analysis of the exposed data after receiving the dataset from Diachenko. The company described the collection as one of the largest known troves of compromised Fortinet-related credentials.

According to Hudson Rock, the dataset contains 73,932 unique firewall URLs across 194 countries and impacts 21,632 unique domains. 

The company says the attackers maintained detailed logs of successful compromises and assembled a database containing verified credentials for organizations across nearly every major industry sector. 

Among the organizations Hudson Rock says appear in the dataset are Foxconn, Samsung, Comcast, Siemens, Lenovo, PwC, Accenture, Oracle, and numerous government agencies and critical infrastructure operators. 

The company also released statistics showing that the highest number of affected devices was in India, the United States, Taiwan, Mexico, Turkey, Thailand, Colombia, Malaysia, Chile, and the United Arab Emirates.

The most common sectors for the listed companies are telecommunications, IT services, financial services, government organizations, healthcare providers, educational institutions, and manufacturing.

One strange aspect of the leak is that many of the exposed credentials were long, complex passwords that would ordinarily be considered difficult to crack.

Believed to be extracted from Fortinet configs

Cybersecurity researcher Kevin Beaumont independently reviewed portions of the exposed data and told BleepingComputer that some of the credentials are authentic.

“I have been able to confirm the authenticity of some of the admin logins and passwords – this looks like a real dump,” Beaumont said.

After further review of the data shared by Hudson Rock, Beaumont published additional findings indicating that the dataset contains credentials for roughly 75,000 Fortinet devices, most of which remain online.

According to Beaumont, the data appears to have originated from exported Fortinet configurations because it contains information, including email addresses, that is typically only accessible through configs.

He also said the affected IP addresses are different from those in the 2025 Belsen Group Fortinet leak, further indicating that this is a more recent and larger collection of compromised devices.

Beaumont said he verified that multiple organizations listed in the dataset were using valid credentials and observed that many affected devices were running relatively recent FortiOS versions.

“The data is legit. It is around 75k devices. Almost all are still online, and Fortinet devices. It appears to be recent data,” Beaumont wrote.

Based on network data from Shodan, Beaumont says the leak contains approximately half of all internet-accessible Fortinet firewalls and said that a majority of the affected devices expose their FortiGate management interfaces directly to the internet.

The source of the configuration data remains unknown, with it unclear whether it was stolen through previously disclosed Fortinet vulnerabilities, a newly discovered flaw, or another method. Neither Diachenko, Hudson Rock, nor Beaumont have identified how the configuration data was originally obtained.

Hudson Rock has created a free FortiBleed lookup tool to check if your organization is impacted.

Organizations in the dataset should immediately rotate passwords associated with Fortinet VPN and administrative interfaces, enforce MFA, examine gateway logs for suspicious activity, and monitor for exposed employee credentials.

BleepingComputer contacted Fortinet regarding the exposed dataset and will update this article if we receive a response.

Security teams log 54% of successful attacks and alert on just 14%. The rest move through your environment unseen.

The Picus whitepaper shows how breach and attack simulation tests your SIEM and EDR rules so threats stop slipping by detection.

Get the whitepaper

During The State of Unreal keynote at Unreal Fest on Wednesday, Epic Games revealed just how it’s embracing generative AI in Unreal Engine (UE). Along with offering the first details on Unreal Engine 6 (UE6), the company discussed new features for Unreal Engine 5.8, which it also released on Wednesday. As part of the latest update, Epic is offering an experimental Model Context Protocol (MCP) plugin that will allow developers to hook gen AI models such as Claude and Gemini into Unreal Engine. It’s looking to make the MCP an integral part of UE6.

Marcus Wassmer, the head of Epic’s development team, wrote in a blog post that the gen AI models can act as “creativity and productivity multipliers so that teams can focus their efforts on the essential creative and technical tasks of development rather than time on time-consuming manual tasks.” 

The blog post went on to state that, “our goal for UE6 is to greatly reduce the tedious work in authoring content to leave more time for creative exploration, and increase the amount of iterations a team can make to polish their content. UE6 will ship with tools and workflows where you can choose to bring your own favorite models, battletested against internal development and in UEFN [Unreal Engine for Fortnite].”

Unreal Engine 5.8 ships today with experimental MCP server support:

Your sources, your pipeline and your workflow—simply configure the MCP plugin and connect to any agent. Get familiar with the MCP server and the PCG Primitive Plugin today and see what teams can build together:… pic.twitter.com/Ca5yZIH443

Advertisement

— Unreal Engine (@UnrealEngine) June 17, 2026

Epic gave a demonstration of Claude Code connecting to UE, then pulling objects from an asset library and placing them in a virtual living room. Developers can still move the objects around manually in the UE editor. 

The company also showed how a developer might use Claude Code in UE to build a city that can be automatically adjusted as assets like parks are added. Along with modifying assets, gen AI models can adjust factors like lighting and match atmospheric conditions to real-world examples.

In a video showing off Unreal Engine 5.8, Epic suggested that developers could use the likes of Claude to “automate asset creation, testing and optimization. The plugin can access core UE systems such as blueprints, assets, levels, materials, meshes and many more.”

It shouldn’t be too much of a surprise that Epic is going all in on gen AI in UE6. Back in November, CEO Tim Sweeney suggested that a “made with AI” tag may be “relevant to art exhibits for authorship disclosure, and to digital content licensing marketplaces where buyers need to understand the rights situation. It makes no sense for game stores, where AI will be involved in nearly all future production.”

In January, the Game Developers Conference published its 2026 State of the Game Industry report, which was based on a survey of more than 2,300 game industry workers. Of those, 36 percent said they were using gen AI tools as part of their job. Most of those using such tools were doing so for research and brainstorming (81 percent) but also for tasks like prototyping (35 percent). However, 52 percent of respondents said they thought gen AI was bad for the industry. That figure was up from 30 percent in the 2025 edition of the survey and 18 percent in 2024. Only seven percent said it was having a positive impact.

Elsewhere at Unreal Fest, it emerged that Epic is merging Unreal Engine 5 and UEFN into a single platform in UE6. One other thing that the company is testing is the ability to pull Fortnite skins into other UE6 games, and to let developers move their skins in the other direction. The company aims to release UE6 in early access in late 2027, with a full release lined up for around 12-18 months later.

Epic had some news to share about collaborations as well. Those creating Fortnite experiences using UEFN will soon be able to make games based on The Simpsons, just as they can currently do with Star Wars IP. The company also revealed that more than 30 gaming collaborations are lined up for Fortnite this year, including Sonic Racing: CrossWorlds, Vampire Survivors, Control Resonant and Phantom Blade Zero.

However, Vampire Survivors developer Poncle appears to have concerns about Epic’s embrace of gen AI. “Following today’s news about gen AI usage by Epic to create all sort [sic] of game assets, including Fortnite characters, we’re currently ‘reviewing’ our collaboration with Fortnite,” Poncle stated on Reddit. “We’ll let you know if anything moves forward.”

Ghana will take on Panama in a must-win contest for both teams as they begin their respective FIFA World Cup 2026 Group L campaigns in Toronto. A defeat would leave either side facing a daunting path, with upcoming matches against European heavyweights England and Croatia.

Quarter-finalists in 2010, the Black Stars enter the match as favorites, particularly with Panama’s star midfielder Adalberto Carrasquilla reportedly nursing an injury and unlikely to be risked. Yet Ghana have had a difficult spell recently, highlighted by their failure to qualify for the most recent 24-team Africa Cup of Nations, for the first time in two decades.