Cloudflare has announced a joint initiative with Mozilla Firefox, Google Chrome, and Microsoft Edge to develop a new internet protocol that verifies whether web traffic is legitimate without tracking users. The protocol, called Private Access Control Tokens, is designed to replace CAPTCHAs and forced logins with anonymous tokens that prove a visitor is human or an authorised bot. Shopify co-developed the technology and the group plans to submit it for formal standardisation.

The announcement comes as bot traffic has officially overtaken human activity online. Cloudflare Radar data shows automated systems now account for roughly 58 percent of HTTP requests to web content worldwide, against 42 percent from people. Cloudflare CEO Matthew Prince shared the milestone on June 3, noting that agentic AI programs browsing on behalf of assistants like ChatGPT and Gemini had accelerated the crossover by about 18 months ahead of his earlier predictions.

PACT works by allowing websites with strong knowledge of a visitor’s identity to issue anonymous tokens. A user’s browser stores the token and can present it to other websites as proof that a real person is behind the session, reducing the need for repeated identity checks. The protocol is designed so that the token cannot be used to track users or reconstruct their browsing history.

“The way we interact with the Internet is facing a fundamental shift,” Cloudflare CTO Dane Knecht said in the announcement. “As AI-powered traffic becomes widespread, existing tools to support its use are too generic and coarse.” He said the collaboration would eliminate the friction caused by security protocols for every visitor, whether human or agent, without sacrificing privacy.

The initiative does not aim to block all automated traffic. Cloudflare has itself embraced agentic AI, cutting 1,100 jobs earlier this year after declaring that AI agents now perform work previously done by humans. For many AI agents there is still a human somewhere in the loop with a legitimate reason to access a website.

PACT is meant to distinguish those authorised agents from malicious scrapers and abuse bots, not to shut down automation entirely.

The browser makers framed the effort as essential to the open web. Bobby Holley, CTO for Firefox at Mozilla, said an “avalanche of automated traffic” was pushing sites toward blunt defences like paywalls, identity checks, and invasive tracking. Erik Anderson, director of engineering for the web platform at Microsoft Edge, called effective privacy-preserving tools critical to combating abuse without unnecessary user friction.

Shopify’s involvement reflects the commercial stakes. Ilya Grigorik, a distinguished engineer at the company, said every extra challenge or false positive in ecommerce can turn a purchase into an abandoned cart. Covert browser fingerprinting and extension scanning have emerged as the default tools for platforms trying to identify users, a practice that privacy advocates and regulators have pushed back against.

PACT would offer a standardised alternative that does not require harvesting device characteristics or tracking browsing behaviour.

The protocol builds on earlier work in the same space. Apple already uses a related system called Privacy Pass, which works with a device’s secure enclave to attest to a user’s identity, and Cloudflare uses Privacy Pass as a signal in its bot management products. The IETF published the Privacy Pass Architecture as RFC 9576, and PACT extends that foundation with broader browser support and a focus on the agentic AI traffic that has reshaped the composition of the web in the past year.

No deployment timeline has been announced. The partners have committed to developing the protocol and submitting it for standardisation, but turning a specification into something that works across billions of browser sessions will take time. Users are already migrating away from platforms that impose AI features without consent, and the question of how to manage automated traffic without alienating human visitors is becoming more urgent by the quarter.

Whether PACT arrives fast enough to matter depends on how quickly the standards process moves and how willing websites are to adopt a system that, by design, gives them less data about their visitors rather than more.

from the setback dept

Well, this is very disappointing. Over the first half of this year, we’ve talked about the resurgence of the Stop Killing Games movement, which aims to push various governments to legislate out the practice of video game publishers sunsetting their games and making them unplayable afterwards. The aims of the movement are simple: publishers can certainly sunset their games that require backend servers to work, but they should make them hostable and playable through fan-run servers if they do, should notify customers well in advance of the sunset date, or should make or alter the games so they can be played independent of the company keeping services running.

To that end, the movement managed to get enough signatures in the EU to get a parliamentary hearing, which was reported to have gone quite well. That’s why it’s a surprising to learn that the EU just ruled out issuing that kind of mandate to publishers. Instead, the EU wants this to be a voluntary process, and what it’s citing as the reason it can’t be done by mandate is breaking my brain.

The European Commission said on Tuesday it cannot require video games to remain playable ​after they are withdrawn from sale, but ‌will work with industry and consumer groups on a voluntary code of conduct for managing games’ “end of life”. The Commission said copyright and other intellectual property rules prevent it from imposing an obligation to keep games playable. It added it would ​work with ​consumer organisations ⁠and authorities to raise awareness of existing rights.

“Active enforcement of these existing consumer ​rights can also incentivise the providers ​to ⁠offer video games with longer lifespans and explore solutions for meeting consumer expectations,” the Commission said in ⁠a ​statement.

Copyright law is my reason why mandates like this should exist. Like, American law, EU copyright law offers protection for a work for the author’s life plus 70 years. After that, the work enters the public domain. Unless, that is, we’re talking about video games that require backend support, in which case it never enters the public domain and instead just vanishes into vapor. And that, I have repeatedly argued, breaks the copyright bargain entirely. In fact, it seems to me that it breaks it so completely that works like that shouldn’t even get copyright protections without rules such as exactly what Stop Killing Games is advocating for.

And this plan for publishers to do all of this voluntarily? Don’t make me laugh. The very gaming companies that the EU wants to take this sort of preservation effort on voluntarily lobbied against codifying preservation efforts. Why would they do that if they were willing to do this all voluntarily?

Finally, the point of all of this is not merely to make games playable for longer. It’s to preserve them as close to infinitely as possible. That should be the aim of any cultural output.

Stop Killing Games hasn’t commented publicly on the decision yet. I doubt the movement will take this defeat lying down, however.

Filed Under: copyright, eu, eu commission, stop killing games, video games

With Microsoft seemingly intent on turning Windows into malware, Macs are increasingly appealing. But while the Apple tax is increasingly diminishing in the consumer laptop space, it’s rife when it comes to compatible monitors.

PC-focused alternatives have different colors, pixel densities and features that rarely play well with Macs, meaning users frequently fight losing battles matching what they see on their MacBook screen with a third-party monitor. But BenQ has taken note.

The JaredFromSubway Ethereum MEV (Maximal Extractable Value) bot suffered a $15 million loss after an attacker manipulated the opportunity-detection logic by creating fake cryptocurrency trading opportunities.

The drain was detected on Saturday by blockchain security firm Blockaid, and today, JaredFromSubway confirmed that the attacker used fake pools and tokens to trick the bot into approving helper contracts.

According to Blockaid, the attacker deployed contracts designed to appear as profitable MEV opportunities to JaredFromSubway’s automated execution system.

The bot automatically analyzed routes and trade opportunities that seemed financially rewarding. It then generated the transactions needed to execute them, granting ERC-20 token approvals to contracts controlled by the attacker.

It appears that the attacker planned the heist carefully, as early transactions served as harmless tests to help confirm the bot’s action routines. Later, the threat actor changed the route so that the allowance was not consumed or revoked after the bot granted approvals.

The attacker accumulated valid spending permissions without immediately using them, reaching up to 92.1614 WETH approved to an attacker-controlled helper contract.

Finally, the attacker used the open approvals to withdraw WETH, USDC, and USDT from the JaredFromSubway MEV bot contract via the transferFrom function.

Karma slaps back

MEV bots are ultra-fast automated trading systems that scan Ethereum and other blockchains for opportunities to make money by exploiting the order and timing of transactions before they are included in a block.

JaredFromSubway is a private MEV operation with no publicly available code, known as one of Ethereum’s most aggressive and visible “sandwich”-bot operations.

In a sandwich attack, the bot detects a user’s pending trade, places a buy order immediately before it, and then sells immediately afterward, profiting from the price movement caused by the victim’s transaction.

The practice is controversial because it often results in worse prices for regular traders while generating profits for the bot operator.

Initially, JaredFromSubway offered a $3 million bounty to the attacker for the full return of the stolen funds, promising no further action would be taken.

After receiving no response, JaredFromSubway increased the bounty to $7.5 million for the return of just 50% of the stolen amount, with $1 million to be given to the community.

JaredFromSubway is also negotiating with “a white-hat hacking group” on the stolen $15 million but there is no confirmation of a deal yet.

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

A short exchange on X pulled antimatter propulsion back into view. Elon Musk posted that in the future a trillion times a trillion dollars will go toward making antimatter so people can travel to other star systems. He added that later civilizations may measure wealth in mass and energy rather than currency. NASA Administrator Jared Isaacman replied that he supports antimatter propulsion.

Antimatter is essentially the mirror image of normal matter. Every electron has a positron, which is identical to it but has the opposite charge. Each proton contains an antiproton. E = mc² states that when particles collide, their mass is converted into energy. When one gram of antimatter collides with one gram of normal matter, it produces energy equivalent to around 43 kilotons of TNT, nearly three times that of the Hiroshima bomb. Chemical rockets can only convert a small portion of their fuel into energy, and nuclear fusion is nowhere near as efficient as antimatter’s near-total conversion.

Sale

Executive Series Apollo 11 Capsule Model – 1:25 Scale, Officially Licensed NASA Spacecraft Replica…

• Officially Licensed NASA Apollo 11 Capsule Replica: Stunning 1:25 scale model of the historic Apollo 11 Command Module, authentically designed with…
• Perfect for Display: Measures 6.25 inches wide x 6 inches high, making it an ideal size for display on shelves, desktops, or in collector’s cases. A…
• Unique Opening/Closing Door Feature: This Apollo 11 model includes a fully functional door that opens and closes, adding realism and enhancing its…

The fact that you can get a lot more oomph from a lot less mass has a significant impact on the type of spacecraft you can build. Current engines must carry the majority of their weight in propellant, which is then discarded as soon as it is spent. An antimatter system might provide a lot more kick for a lot less weight. What used to take six to nine months to get to Mars could now be done in weeks, and getting to the nearest stars, which takes tens of thousands of years with current technology, could be done in a human lifetime or a few decades at a few percent of lightspeed. Less time in space implies less radiation and weightlessness for any crew, which is a significant benefit.

Several ideas have been proposed to turn annihlation into thrust. One approach is to combine streams of antimatter and matter in a dedicated chamber, where the resultant particles and radiation are blasted out the rear. Another does the same thing, but uses the energy to heat up propellant such as hydrogen, which then expands and generates thrust. Then there’s the concept of using tiny amounts of antimatter to initiate larger nuclear fusion or fission reactions, which would stretch the limited antimatter supply even further. Every one of these approaches is trying to solve the same problem: how to take the energy release and turn it into push without wiping out the ship.

The trouble is that creating antimatter is an absolute nightmare. Particle accelerators generate it by smashing particles together, however the method creates extremely little amounts. After years of experimentation, places such as CERN have only been able to produce nanograms. Getting to the amounts required to create even a small probe would necessitate discovering new ways to manufacture it and much more effective ways to collect it.
[Source]

No matter your lifestyle or budget, Samsung has a range of phones to fit what you’re looking for. From baseline options such as the Galaxy S25 and S25 Plus (or the newer S26 and S26 Plus) to more premium picks like the Galaxy S26 Ultra, you can get solid performance, good cameras and long-lasting batteries regardless of how much you spend. Recently, the Galaxy S26 Ultra received a CNET Lab Award for having the fastest wired charging of 33 phones we tested. For a novel design, you can choose Samsung’s pricier Galaxy Z Flip 7 or Z Fold 7, or even the super-slim Galaxy S25 Edge. And for a more affordable option, Samsung’s A26, A36 or A56 might be a good fit. Our roundup can help you find the phone that best fits your needs.

What is the best Samsung phone?

The Samsung Galaxy S26 Ultra — priced at $1,300 — comes packed with maximum features that, for most people, is more than necessary. For the rest of us, last year’s $800 Galaxy S25 is a standout among its Galaxy counterparts. Even with the release of the S26, the S25 remains a solid pick because it has much of the same hardware, software and AI capabilities as it pricier (and newer) peers.

The Galaxy S25 has a very capable triple rear camera setup that is versatile in capturing bright outdoor scenes and candid moments inside under mixed lighting. Overall, the Galaxy S26 Ultra is ideal for Android fans who prioritize fast performance, versatile cameras and a spacious screen. 

How to pick the right Samsung phone for you

Deciding which Samsung phone is right for you comes down to what you want in a phone and how much you’re willing to spend. If you want the largest screen available on a standard Samsung phone, enjoy note-taking with a stylus and need a camera with a significantly closer zoom, the Galaxy S26 Ultra is the right choice. You’ll also have to spend $1,300 unless you score a trade-in deal. 

Those who don’t need the stylus, prefer a more compact phone and still want a good camera should consider the Galaxy S25 or Galaxy S25 Plus. If you just want the basics, like a spacious screen, 5G and a decent camera, consider the Galaxy S25 FE. Those looking for the flashiest tech around — and who also have deep pockets — should consider the Galaxy Z Fold 7, Z Flip 7 or even the slightly cheaper Z Fold 7 FE. 

Finding the best Samsung phone will ultimately come down to preference. Choosing among so many options can get complicated, so here’s how to decide which Samsung phone is best for you. Refer to our phone buying guide for more tips on how to choose the right device.

Best Samsung phones in 2026

Advertisement

We test every phone in real-world scenarios focusing on its features, design, performance, cameras, battery life and overall value. We document our findings in an initial review that is periodically updated when there are new software updates or to compare against new phones from competitors like Apple, Google, OnePlus and Samsung.

Photography

Photography is a major focus for most phones these days, so we take pictures and videos of various subjects in a variety of settings and lighting levels. We try out any new camera modes such as 4K 120fps slow motion video that debuted with the iPhone 16 Pro and 16 Pro Max or AI reframe and focus on the Motorola Razr Plus (2024).

Advertisement

Battery life

Battery testing is conducted in a variety of ways. We assess how long a phone lasts during a typical day of use, and note how it performs during more focused sessions of video calls, media streaming and gaming. We also conduct a video playback test, which isn’t always included in the initial review and is added later in an update.

Performance

We perform processor-heavy tasks like editing photos, exporting videos and playing games. We evaluate whether a newer version of a particular phone includes enough features to make it worth upgrading from older models.

We use benchmarking apps to measure the performance, alongside our own anecdotal experiences using the phone for our review. Of particular note are how graphics and animations look. Are they smooth? Or do they lag or stutter? We also look at how quickly the phone switches between horizontal and vertical orientations and how fast the camera app opens and is ready to take a photo.

Read more: How We Test Phones

Looking for the most recent Mini Crossword answer? Click here for today’s Mini Crossword hints, as well as our daily answers and hints for The New York Times Wordle, Strands, Connections and Connections: Sports Edition puzzles.

Need some help with today’s Mini Crossword? It wasn’t too tricky today, but read on for all the answers. And if you could use some hints and guidance for daily solving, check out our Mini Crossword tips.

If you’re looking for today’s Wordle, Connections, Connections: Sports Edition and Strands answers, you can visit CNET’s NYT puzzle hints page.

Read more: Tips and Tricks for Solving The New York Times Mini Crossword

Let’s get to those Mini Crossword clues and answers.

Mini across clues and answers

1A clue: Baseball officials
Answer: UMPS

5A clue: Singer/songwriter Kahan with the 2026 #1 album “The Great Divide”
Answer: NOAH

6A clue: Shade of orangeish pink
Answer: CORAL

8A clue: Voice of the smartphone generation?
Answer: SIRI

9A clue: Partner channel of ABC
Answer: ESPN

Mini down clues and answers

1D clue: Someone giving off old guy vibes, in Gen Z slang
Answer: UNC

2D clue: Tallest animal in North America
Answer: MOOSE

3D clue: Setting for “Ratatouille”
Answer: PARIS

4D clue: #, in music
Answer: SHARP

7D clue: ___-Manuel Miranda
Answer: LIN

Not every company can or should build their own frontier AI language model. However, the harness controlling the model is something that most enterprises can and should customize for their specific purposes.

Of course, this is easier said than done. Agent harnesses are still largely tuned through manual, ad hoc debugging — a process that relies heavily on intuition rather than systematic feedback loops, making it difficult to keep pace with rapidly evolving LLMs.

To solve this challenge, researchers at the Shanghai Artificial Intelligence Laboratory have introduced “Self-Harness,” a new paradigm in which an LLM-based agent systematically improves its own operating rules. By examining its own execution traces to apply edits, the system trades manual guesswork for empirical evidence.

Self-improving harnesses can enable development teams to deploy robust custom agents that continually adapt their own execution protocols to overcome model-specific weaknesses.

The challenge of harness engineering

An LLM-based agent’s performance is not determined solely by its underlying base model, but also by its harness: the surrounding system that provides context and enables the model to interact with the environment. A harness includes components like system prompts, tools, memory, verification rules, runtime policies, orchestration logic, and failure-recovery procedures.

This layer is crucial because many common agent failures stem from the harness rather than the model. For example, an agent may report success without checking the model’s response (e.g., running the code to see if it passes the tests), or it might retry a failed action repeatedly. The harness is also responsible for preventing context rot or overload when the agent’s interaction history grows very large. Examples of popular harnesses include SWE-agent, Claude Code, Codex, and OpenHands.

Harness engineering remains a significant challenge, but the bottleneck isn’t necessarily that humans are too slow or incapable.

In fact, Hangfan Zhang, lead author of the Self-Harness paper, told VentureBeat that “in many cases, an experienced engineer with deep domain knowledge can still propose better changes than an LLM can today.”

Instead, the true bottleneck of manual engineering is that it relies heavily on ad hoc debugging rather than a verifiable, empirical feedback loop. “The deeper issue is that the current harness-engineering paradigm often lacks a systematic feedback loop,” Zhang explained. “Many edits are made based on intuition, a few observed failures, or ad hoc debugging.”

With new models being released at a rapid pace, depending on human intuition to manually tune model-specific harnesses becomes increasingly costly and untenable. While some approaches use stronger models to improve the harnesses of weaker target agents, this dependence on external guidance has its own challenges, as these models may be costly, unavailable for frontier models, or mismatched to the target model’s failure modes.

How Self-Harness works

The Self-Harness paradigm enables an LLM-based agent to improve its own harness without relying on human engineers or stronger external models.

This continuous self-evolution is driven by a three-stage iterative loop that turns behavioral evidence into harness updates:

• Weakness mining: Starting from an initial harness, the agent runs a set of tasks, producing execution traces with verifiable outcomes. The agent categorizes failed traces and tries to detect model-specific failure patterns.

• Harness proposal: Based on these failure patterns, the agent uses a “proposer” role to generate a set of diverse yet minimal harness modifications, each tied to a specific failure mechanism to avoid overly general corrections.

• Proposal validation: The system evaluates candidate modifications through regression tests. An edit is promoted only if it improves performance without causing measurable degradation on held-out tasks. If multiple candidate modifications pass the regression tests, they are merged into the next version of the harness, which then serves as the starting point for the next iteration.

To visualize why an enterprise would need this, imagine an automated issue-fixing agent that reads internal documentation, writes patches, and opens pull requests. If the company updates its documentation style, the agent might suddenly fail, pulling the wrong context or writing bad patches.

On the surface, the agent simply looks broken. But Self-Harness turns this ambiguous failure into a solvable problem. “The failure traces expose where the agent is misusing the new documentation format; the proposer can generate a targeted harness edit… and the evaluator can decide whether that edit improves the failing cases without regressing other cases,” Zhang said.

Self-Harness in action

The researchers evaluated Self-Harness on Terminal-Bench-2.0, a benchmark that tests general tool-based execution, including artifact management, command use, verification behavior, and recovery from execution errors. They applied Self-Harness with MiniMax M2.5, Qwen3.5-35B-A3B, and GLM-5.

To isolate the impact of the self-evolving harness, they started with a minimal harness built upon the DeepAgent SDK, containing only the benchmark-facing system prompt, and the default filesystem and shell tools. The model backend, tool set, benchmark environment, and evaluator were kept unchanged while only the harness was allowed to vary.

The quantitative results show that agents improved their performance through automated harness edits. On held-out tasks, performance jumped significantly across the board, ranging from 33 to 60 percent relative improvements for different models.

Importantly, an explicit acceptance rule promotes only those edits that improve performance without introducing unacceptable regressions. What makes Self-Harness powerful for enterprise applications is that it doesn’t simply make the prompt longer or add generic instructions. Instead, it introduces targeted changes that reflect the recurring problems each model encounters during execution.

For example, under the baseline harness, MiniMax M2.5 would get stuck endlessly exploring dataset configurations until the execution environment timed out, failing to produce any deliverables. Through Self-Harness, the system identified this specific flaw and wrote a “loop breaker” into its runtime policy, forcing the agent to stop and redirect its approach after 50 tool calls. It also added a rule to create an initial version of required artifacts as early as possible.

On the other hand, Qwen-3.5 had a habit of hitting a file overwrite error and then blindly retrying the same command repeatedly, eventually deleting necessary files out of confusion before stopping. The self-harness fixed this by introducing a strict command-retry discipline (forbidding exact duplicate commands) and a mechanism that forced the agent to immediately recreate any missing artifacts if a file error occurred.

GLM-5 struggled to preserve environment changes across different commands, and would often waste time on massive downloads or finalize tasks even when sanity checks were failing. Its self-generated harness introduced rules instructing the agent to persist PATH variables across shell sessions, limit external compute, and repair any failed sanity checks before concluding its run.

The hidden costs of automated harnesses

While Self-Harness automates the tedious work of tracking down idiosyncratic model failures, decision-makers must be realistic about the trade-offs. Replacing human engineering with automated trial-and-error requires significant computational overhead.

“Self-Harness replaces part of the human engineering burden with repeated proposal generation, parallel candidate evaluation, and regression testing,” Zhang said. “That can mean more API tokens, more latency during optimization, and more infrastructure for running evaluation tasks.”

Also, this system relies on the accuracy of its evaluation pipeline. During their experiments on Terminal-Bench-2.0, the researchers relied on strict, deterministic verifiers to ensure the agent’s edits were actually helpful. Without this rigorous ground truth, an automated system risks promoting bad updates. “[The] evaluation system is not an optional component; it is what lets us trade human intuition for empirical evidence,” Zhang said.

This reliance on strict verifiers also dictates where Self-Harness should be deployed. “The best deployment targets today are environments where failures can be measured and where trial-and-error is relatively safe,” Zhang said, pointing to coding, internal workflow automation, and DevOps data pipelines as ideal use cases.

Conversely, enterprises should avoid fully automating harnesses in high-stakes or subjective fields. “The clearest red flags are domains where evaluation is subjective, delayed, non-deterministic, or costly to get wrong, such as medical decision-making, safety-critical infrastructure, or legal decisions.”

From prompt tweakers to feedback architects

The introduction of self-improving agents does not mean coding or enterprise workflows will suddenly become human-free. The quality of collaboration between the human engineer and the AI is still paramount and difficult to capture with automated benchmarks. 

Instead, the engineering profession is moving up the abstraction layer. “The role of enterprise engineers will shift from manually patching individual prompts or tool calls toward designing the feedback systems that make agent improvement possible,” Zhang predicted. Moving forward, “the engineer becomes less of a prompt tweaker and more of a feedback architect.”

As foundational models grow more capable, they will naturally absorb many capabilities that currently require manual harness engineering. “But once that happens, the harness will not disappear; its scope will move outward to connect the model to richer external environments,” Zhang said. “Until that boundary moves beyond what humans can evaluate, humans will remain critical providers of feedback.”