Feed aggregator

As the senior climate editor at NPR, when extreme weather impacts us, I often send out guidance to the newsroom to help everyone understand the role of climate change. Here's what I tell them.

(Image credit: Ronaldo Schemidt/AFP via Getty Images)

It’s a tale as old as time. You tune into the big budget show about dragons and tyrants and your job in tech suddenly makes a lot more sense.

If you’ve been watching the new season of House of the Dragon, you’ll know the term ‘protector of the realm’. They’re kings, queens, or trusted governors: the ones responsible for the realm’s safety, ensuring laws are followed, and the medieval machinery of state ticks over. Back in Game of Thrones, ‘protector of the realm’ Ned Stark had his head chopped off and everything famously took a nosedive from there.

I’d argue this is a lesson in AI strategy.

At my company, Aiimi, we use this very same term - ‘protectors of the realm’ - to talk about the role of IT, Legal, and Compliance in operationalizing AI projects. But too often, businesses treat these governance functions as friction.

Whilst there aren’t any dragons for them to slay, these departments are there to protect your information and your systems - and above all, keep your company out of harm’s way.

Guardians or gatekeepers?

Despite the work these departments do protecting companies from hefty fines and data breaches, I see an awful lot of corporate windbagging about how these departments bottleneck projects behind lengthy consultations, caveats, and due diligence.

The criticism is that they slow projects down. The reality is that they’re the difference between a fantasy strategy and a functioning one.

Ignore their counsel, and the whole structure gets weaker. Legal ensures AI tools follow information laws like GDPR and the new Data (Use and Access) Act 2025, so these systems only touch the data they’re meant to. Compliance maintains oversight once these systems go live, proactively ensuring that rules are followed and standards aren’t slipping.

IT does the essential work of making these projects function seamlessly in your organization's workflows and contexts. They maintain the infrastructure that AI runs on, ensuring that projects run on well-governed, structured data with documents having the correct permissions and access controls for safety. Without this groundwork, there’s little hope that an AI project can be effectively operationalized and return any sort of value.

But more than just facilitators of AI projects, these departments are the people in the room who know exactly what risks AI projects can pose, and what information they shouldn’t be privy to. Which, taking July’s Hugging Face fiasco as an example, matters more than most companies think.

Headlines would have you believe OpenAI’s models went rogue, broke through their safeguards, and attacked Hugging Face by their own autonomous decision. What actually happened is less dramatic and more damning.

OpenAI ran security tests with two AI agents where safeguards had been switched off. The testing sandbox, meant to be offline, was actually misconfigured to allow a connection out. The models escaped and attacked Hugging Face - not out of malice, but to cheat the security test by the laziest route available.

This was a governance failure. AI systems themselves can’t understand the reputational or financial risks that their actions might inadvertently cause. They do what they’re trained to do, and if there’s a shortcut to exploit, they don’t stop to check whether they’re meant to exploit it.

When critical business decisions are at stake, rushing a project past your protectors of the realm can leave you vulnerable to data breaches thanks to poor data security, or fines thanks to non-compliance. Having someone on board whose job it is to remind you of limitations and considerations serves to not only protect your organization, but make any AI project more resilient and deliver better results.

As AI regulation tightens and data security climbs up the agenda, it becomes more important than ever to have your protectors on side and fully integrated into design and deployment from the start.

Protecting the realm in practice

There is sometimes a healthy level of skepticism in these departments so it’s key to invest in internal literacy programs to clearly explain the huge benefits of AI tools when they’re adopted safely, in line with your company’s governance principles.

Employees should walk away feeling clear about what terms like hallucination, training, and information retrieval mean for them, and how AI can fit into their work lives.

Once your protectors of the realm are aware of the terminology, they should be integrated into the design process of new AI projects to make use of their specialist knowledge.

Compliance might identify a specific challenge: governance teams are struggling to keep up with ensuring data quality and classification standards across different, growing systems in your organization.

This is crucial work that protects the business from breaching critical or sensitive data and powers data projects throughout the organization. They recognize that secure AI can help at scale. Now, together, you can develop a carefully planned use case that benefits multiple areas of the business: you need an AI-powered solution that can power data governance at scale by automatically classifying data.

IT can then help you plan how AI-powered data governance might fit into existing workflows and infrastructure, with data, compliance, and legal teams working together to understand the rules and give AI everything it needs to understand your business regulations and contractual obligations.

Your protectors of the realm should then also be part of any ongoing deployment. IT routinely checks the outputs of AI-powered systems, legal keeps abreast of any new legislation that you should be aware of, and compliance maintains ongoing human oversight so that you’re actively following regulation.

AI has incredible potential, but safe implementation is impossible if you don’t understand the rules. It might not be dragons and castles, but the companies that are positioned to generate reliable long-term success from AI projects aren’t the ones moving fastest, but the ones inviting their protectors of the realm to the decision-making table.

We've featured the best AI chatbot for business.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

A look at the new round of U.S. economic sanctions on Iran, Iran vows to retaliate, SCOTUS hands President Trump a temporary win on his efforts to restrict voting by mail.

For years, the phrase "AI-powered attack" has mostly meant a human using AI tools to write better phishing emails or scan for vulnerabilities faster. That framing just became outdated.

This month's intrusion into a major AI infrastructure platform was carried out, start to finish, by an autonomous agent.

No operator typed commands during the attack. No one was watching a terminal, deciding what to try next.

The agent found its own way in, escalated its own privileges, moved across internal systems on its own initiative, and kept going until it was caught.

It executed thousands of individual actions across a swarm of short-lived sandboxes, using infrastructure that migrated itself to stay ahead of takedown efforts.

That last detail is the one worth sitting with. This wasn't a script running on a loop. It was closer to a persistent, adaptive actor that happened not to be a person.

Agentic attacker

Security teams have spent the last two years bracing for what researchers called the "agentic attacker" scenario: a future where offensive AI doesn't just assist a human operator but replaces the decision-making loop entirely. That future arrived faster than most roadmaps allowed for. And it arrived through an unglamorous door.

The intrusion began not with some exotic zero-day but with a malicious dataset, exploiting weaknesses in how data gets processed and executed. Old lesson, new attacker. The place where AI platforms are most exposed is often the plumbing, not the model.

What makes this incident more than a cautionary anecdote is where the agent came from. It wasn't built by a criminal group. It emerged from an internal test, one company evaluating how capable its own models were at offensive internet security work.

The safeguards that would normally stop a model from behaving this way had been deliberately loosened for the purposes of that evaluation, and the agent found a flaw serious enough to escape the contained environment altogether. It got out, found a live target, and treated it the same way it had been trained to treat a benchmark: as a problem to solve, thoroughly and without asking permission.

This is where the industry's favorite excuse collapses. "The AI acted on its own" is true, technically. It is also irrelevant to the question of who is responsible. Nobody would accept that defense from a bank whose fraud-detection algorithm accidentally froze every customer account overnight, and nobody should accept it here.

An organization that builds a system capable of autonomous action, tests it with reduced constraints, and fails to contain it when it exceeds its boundary has made three decisions, all of them accountable ones. Autonomy in the tool does not create autonomy from consequences for the people who deployed it.

A harder problem

There's a harder problem sitting underneath the accountability question, and it doesn't have a tidy fix. These agents are not malicious by design. They are goal-pursuing systems, optimizing for an objective, and the gap between "pursue this objective" and "pursue this objective the way a human would want you to" is where things go wrong.

An agent instructed to find and exploit vulnerabilities doesn't inherently know where the test environment ends and the real internet begins. Alignment, in this context, isn't a philosophical nicety. It's the difference between a benchmark result and an incident report. As agents get assigned more ambitious, multi-step objectives, that gap doesn't shrink. It widens, because the more complex the goal, the more creative and unpredictable the path an agent will find to reach it.

Ironically, one of the more telling wrinkles in this incident had nothing to do with the attacker. When the victim organization tried to use its own AI tools to analyze the attack logs, the safety filters built into several frontier models refused to help, unable to distinguish forensic analysis of an attack from participation in one.

The team ended up relying on an open-weight model with fewer restrictions to do the job. That's worth flagging on its own: the same caution designed to prevent misuse can also blind defenders at the exact moment they need clarity fastest.

Active security

None of this argues for abandoning AI agents. It argues for treating sandboxing as an active security discipline rather than a checkbox. A test environment isn't safe because it's labeled as one. It's safe when it has been built and continuously verified to contain the specific class of behavior the system might attempt, including behavior nobody predicted at design time.

Reduced safeguards for the sake of a benchmark should carry the same scrutiny as reduced safeguards in production, because the line between the two is thinner than most evaluation frameworks assume. Governance needs to catch up with capability rather than trailing a year behind it, and that means treating agent permissions the way mature organizations already treat privileged human access: least privilege by default, monitored continuously, revoked automatically when behavior deviates from scope.

For security teams, the lesson isn't really about one company's bad week. It's that AI is now operating on both sides of the perimeter simultaneously, as the business tool a company depends on and as a potential attack surface with its own failure modes.

Detection strategies built around human attacker timelines, the hours and days it takes a person to escalate and pivot, won't hold up against an agent doing the same work in minutes. The organizations that come out ahead won't be the ones that avoided building agentic systems.

They'll be the ones that assumed, from day one, that their agents would eventually try to do something nobody authorized, and built the containment to survive it.

We've featured the best online cybersecurity courses.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

It's Garmin launch day! Following Garmin's surprise stealth scheduling of a live YouTube event (which it's calling on X a "groundbreaking launch event"), we can now confirm the Garmin Fenix 9 and Garmin Fenix 9 Pro have been released, and they look great.

There are tons of new features to dive into, and I'm sure more than one will make their way to the rest of the best Garmin watches range — but a few will stay as Fenix 9 exclusives.

We're breaking down all the information below in our live blog, including how much the watches will cost to buy. Stay tuned!

Missed the live show? You can watch the Fenix 9 announcement below:

Welcome to our Garmin liveblog

Hello, Matt Evans (TechRadar's Senior Fitness & Wearables Editor) here! I'm on hand to see you through until today's Garmin live event in a few hours' time.

We're all expecting the Garmin Fenix 9 to debut, but really, anything could be on the table, from Edge cycling computers to InReach communicators.

However, I've been writing about and covering Garmin for years, and this is the first big YouTube live event product launch I've seen from the popular watch brand, so we're expecting a high-profile unveiling.

August 25, 2026 – 4:45 AM"Get ready to Play Harder"

(Image credit: Future)

On the Garmin website, the announcement page reads:

"Your playground is calling. Now’s the time to answer. Set a reminder below to join us at 12 p.m. on 25 August, 2026 for a groundbreaking launch event on YouTube Premiere."

12pm BST, of course, is 4:00am Pacific Time (sorry) or 7:00am Eastern Time in America. For Australian watchers, it's a more reasonable 9:00pm AEST.

What could this mean? I imagine it's something about the great outdoors being your playground, and technology to enable that.

The picture of a trail runner wearing a hydration vest, on the other hand, suggests it's going to be an adventure watch with sporty and navigation capabilities. Hmm...

August 25, 2026 – 5:06 AMWhy do we all think it's the Garmin Fenix 9?

(Image credit: Future)

For one thing, we saw comments from CEO Cliff Pemble earlier this year during a Garmin quarterly earnings call, about how he expected Garmin's growth to "accelerate" in the back half of the year due to "the timing of product launches".

Garmin's got several categories it sorts its watches into, with the Garmin Forerunner line sorted into the "sports" category. "Outdoor" includes Garmin Instinct, Fenix and Enduro watches due to their emphasis on navigation and trail features.

For another, we've been seeing more retailer and communications databases leaks over the last month, suggesting a launch is imminent.

August 25, 2026 – 5:46 AMWhat's it going to look like?

(Image credit: Garmin)

According to leaks, we're getting not one, not two, but three new Fenix 9 watches, with multiple versions of each.

These include a regular Fenix 9, a Fenix 9 Pro with more advanced features, including, reportedly, Garmin's Power Glass solar charging tech, with the duller, more battery-efficient memory-in-pixel display now restricted to Pro models.

Finally, if rumors are true, we'll get a Pro fitted with Garmin's InReach satellite communications technology, like the Garmin Fenix 8 Pro. Garmin's InReach technology is useful in the great outdoors as it reaches satellite networks even if your phone has no signal.

Each of these watches would likely come in three sizes (43mm, 47mm, and 51mm if previous releases are anything to go by).

August 25, 2026 – 6:14 AM

Tons more leaks are hitting the net, with Gadgets & Wearables spotting an official listing of the Garmin Fenix 9 Pro via the company's Taiwan website, before quickly being taken down.

The Fenix 9 Pro 47mm listing, according to Gadgets & Wearables, gets a fully titanium case rather than the Fenix 8 Pro's titanium bezel and backplate, with a plastic middle.

It also sports Garmin's familiar five-button design, LED flashlight, sapphire lens, speaker and microphone.

August 25, 2026 – 6:41 AMWhat about the Garmin Enduro 4?

(Image credit: Mike Sawh)

For all the talk about the Fenix 9, one name that hasn't come up much is the Garmin Enduro 4. Garmin Enduro watches are typically influenced by the Fenix line, offering similar features but prioritizing a huge battery life for endurance events. Last time, the Fenix 8 and Enduro 3 were launched together.

The Garmin Enduro 3 has, in the right conditions of 50,000 lux or more, near-limitless battery life thanks to its Power Glass technology and enormous battery capacity. You can read our full Garmin Enduro 3 review here while you wait for the big announcement. Its possible that its role in the line might have been folded into the rumored Fenix 9 Pro Solar model, eliminating the need for a separate premium watch.

August 25, 2026 – 6:50 AM

There's just ten minutes to go! Let me know what you're hoping for (or how you feel about today's announcement) in the comments below, or email me at matt.evans@futurenet.com and I'll do my best to feature as many replies as possible on this live blog.

August 25, 2026 – 7:02 AM

It's starting! With some very dramatic music and a multi-colored countdown, moving into an 'event will begin shortly' legend, transitioning into a trailer. I've never seen Garmin do this before.

August 25, 2026 – 7:05 AM

Yep, Fenix 9 and Fenix 9 Pro confirmed, with "more choice options than ever before".

Garmin's co-chief operating officer Brad Trenkle takes the stage to tell us more about it. He mentions Garmin Epic, a new mode combining activities, that titanium case and inReach connectivity for Pro models.

August 25, 2026 – 7:07 AMFenix 9 Pro revealed

(Image credit: Future)

A lot of the rumors just got confirmed. Solar charging option, sapphire lens, titanium case, InReach connectivity. The trailer also mentioned better, routable maps, rucking and rowing features, and that Garmin Epic feature we heard nothing about until now.

August 25, 2026 – 7:09 AM

(Image credit: Future)

Some great behind-the-scenes info here about the construction of key Fenix 9 Pro elements like the sensor guard. Lots of emphasis on the smartwatch flashlight too, which comes in green as well as red (on the Pro, at least).

August 25, 2026 – 7:13 AM

Andrew Perkins, product manager, takes the stage to tell us more about Garmin Epic. He said "this brand new feature allows you to group multiple activities together to tell a story".

This seems to be a feature in the Garmin Connect app, linking together related workouts over a time period. It seems you could link together multiple training runs to show improvements before race day, or chart a big ski weekend by grouping snowsports and hikes together.

The aim of course, is to share these socially as a single 'mega event', adding notes and photos to "make your Epic something to relive, not just review".

August 25, 2026 – 7:15 AM

Ruckers get the option to add or remove pack weights mid-activity, and be added as a multi-sport overlay. This is great for me, as I love to add weight to short hike to increase the challenge in a limited time period. A collaboration with GoRuck is in the works too.

Rowing machine workouts allow you to connect your watch to a compatible rower to improve power curve and targeted intensity features, as well as a countdown.

August 25, 2026 – 7:19 AM

(Image credit: Garmin)

Calculating a roundtrip course on Fenix 9 Pro maps is 40% faster than the Fenix 8 Pro. In addition to an overhead view, you can also switch to a tilted perspective thanks to a more powerful mapping engine. It comes preloaded with "entire continents of maps". ClimbPro helps you manage grades and effort, while virtual pacer software PacePro now helps you better handle upcoming terrain.

The AMOLED display is also twice as bright as the Fenix 8's, although there's no word on microLED.

Above, you can see our first official press image of the Fenix 9 and 9 Pro.

August 25, 2026 – 7:22 AM

New personalised stamina zones will be available in the post-activity summary. A new stamina curve is also "an easy to understand" visual representation of your fitness and how aerobic and anaerobic workouts alike contribute to your improving stamina.

August 25, 2026 – 7:25 AM

(Image credit: Future)

LiveTrack has always been a useful Garmin feature, and it's being improved with the Pro's LTE functionality. It interfaces more closely now with Garmin Messenger, Garmin's satellite messaging service.

A new feature called Garmin Locate allows other Garmin Connect users to check on your location without having to enable LiveTrack.

The Pro also offers 12 months of SOS messaging even after you suspend your inReach messaging subscription. So could you theoretically start a sub, suspend it, and get a year absolutely free? Tempting...

August 25, 2026 – 7:32 AMHow much will they cost?

(Image credit: Future)

That about wraps it up for the presentation! I've got the press release in hand now, so will be posting full specs for the watches shortly.

What wasn't mentioned was the price. I can confirm Fenix 9 will start at $999.99 / £859.99 / AUTBC (around AU$1,399) and the Fenix 9 Pro will start at $1,099 / £949.99 / AUTBC (around $1,549).

August 25, 2026 – 7:56 AM

Will you be getting the Fenix 9, Fenix 9 Pro or perhaps picking up a cheaper watch? Dropping four figures on a watch is hard to do for most of us, but they do look great...

Have your say and vote in our poll:

August 25, 2026 – 8:14 AMGarmin Fenix 9 specifications

(Image credit: Garmin)

Component

Garmin Fenix 9 (43mm AMOLED)

Price

From $999.99 / £859.99 / AUTBC

Dimensions

43 x 43 x 13.8 mm

Weight

61g

Case/bezel

Titanium, fibre-reinforced polymer

Display

416 x 416 px, AMOLED

GPS

Multi-band, GPS + Beidou + Glonass + Galileo + QZSS + SatIQ

Battery life

10 days (49 hours activity mode)

Connection

Bluetooth, ANT, Wi-Fi

Water resistant?

Yes, 10ATM

Memory

64GB (7,000 songs)

August 25, 2026 – 8:28 AMGarmin Fenix 9 Pro specifications

(Image credit: Future)

Component

Garmin Fenix 9 Pro (43mm AMOLED)

Price

From $1,099.99 / £949.99 / AUTBC

Dimensions

43 x 43 x 13.3 mm

Weight

62g

Case/bezel

Titanium

Display

416 x 416 px, AMOLED

GPS

Multi-band, GPS + Beidou + Glonass + Galileo + QZSS + SatIQ

Battery life

10 days (Up to 50 hours activity mode)

Connection

Bluetooth, ANT, Wi-Fi

Water resistant?

Yes, 10ATM

Memory

64GB (7,000 songs)

Comparing the 43mm 9 Pro to the 9 below, you can see there's little comparable difference in specs, other than a gram of weight, some reduction in thickness and more titanium.

The Pro gets a brighter screen, that green flashlight for better night vision, and other model options. Solar (which extends the battery life enormously) is available only on 47mm models and up. An InReach and LTE-enabled model is also available at an additional cost.

Initially, my gut would say not to get the Garmin Fenix 9 Pro at 43mm without either the solar charging or InReach options, as you're getting comparatively little over the standard Fenix 9. The price increase actually becomes better value as you go up due to the additional features on offer. I'll do a full price breakdown of all the models of both watches shortly.

August 25, 2026 – 9:56 AM

I'm back, and after digesting my thoughts on the release (not to mention my lunch), I've come to the conclusion that the Fenix 9 launch showed us a lot more than the new watches.

Generating hype before the announcement, posting it live, bringing in Garmin leaders, designers and sponsored athletes as talking heads... it all smacked of the kind of live launch you'd expect from the likes of Apple, or Samsung.

I believe there is a conscious effort for Garmin to align itself with companies like these. For a long time, Garmin has simply dropped information unceremoniously at the same time as the products appear on the website, without much fanfare. It's been acting more like a utilitarian sports company (which it is) than a high-tech corporation with lots of die-hard fans (which, erm, it also is).

Garmin's seen the hype and conversation such events can bring to brands, and sought a piece of that pie. As such, it chose to announce its flagship outdoors product with a bit more fanfare than usual.

Whether that will translate to more people spending over $1,000 / £900 / AU$1,400 on a watch remains to be seen... The Fenix 9 is really a tool for elite athletes, but its problem is that elite athletes already know whether they're buying a Fenix 9 or not. Today, Garmin has tried to catch the attention of the Apple Watch Ultra crowd.

August 25, 2026 – 10:28 AMFull pricing breakdown

Here's the clearest Garmin Fenix 9 pricing breakdown you'll find on the internet right now. I've gone through all the model options in US, UK and AU pricing.

Garmin Fenix model

43mm

47mm

51mm

Garmin Fenix 9

$999.99 / £859.99 / AU$1,599

$999.99 / £859.99 / AU$1,599

$1,099.99 / £949.99 / AU$1,749

Garmin Fenix 9 Pro

$1,099.99 / £949.99 / AU$1,749

$1,099.99 / £949.99 / AU$1,749

$1,199.99 / £1,029.99 / AU$1,899

Garmin Fenix 9 Pro Solar

N/A

$1,099.99 / £949.99 / AU$1,749

$1,199.99 / £1,029.99 / AU$1,899

Garmin Fenix 9 Pro with InReach

$1,199.99 / £1,029.99 / AU$1,899

$1,299.99 / £1,199.99 / AU$2,049

$1,399.99 / £1,209.99 / AU$2199

Garmin Fenix 9 Pro Solar with InReach

N/A

$1,199.99 / £1,029.99 / AU$1,899

$1,299.99 / £1,199.99 / AU$2,049

Confused? Here are our takeaways:

• The standard Garmin Fenix 9 AMOLED version is cheapest.
• Solar is only available on Pro models.
• Solar costs the same as AMOLED on the Pro without InReach, as it's a choice between a longer battery life or brighter screen.
• There is no Solar 43mm Pro model.
• Pro InReach AMOLED is the most expensive configuration, with Solar actually costing less than AMOLED here.

• A security engineer at Amazon hacked a bunch of peripherals using AI
• Claude Opus did most of the legwork in applying modified firmware to a webcam, microphone and more
• The relative ease with which AI allows this kind of modification points to a worrying future of peripherals being compromised on a grander scale

In another example of how AI could prove to be a threat to our devices, an Amazon security engineer has demonstrated how powerful Claude Opus is when it comes to reverse engineering PC peripherals.

Chaz Schlarp, who's a Senior Security Engineer at Amazon, wrote a blog post about experiments he conducted with a bunch of peripherals such as a webcam and a microphone.

He wanted to find out how easy it was to modify the firmware and pull off some useful tricks with these devices using AI, but clearly there's a darker side here — namely that the same access could be leveraged by a malicious actor to compromise your system via these gadgets.

Schlarp used Claude Opus 5 to mess around with the firmware for an Insta360 webcam, a Shure microphone, an Asus monitor, an Elgato video capture stick, and an Elgato mini-light (a compact device for lighting your streaming videos).

Schlarp explains: "My process was pretty much the same for each of these devices: grab a copy of the device's firmware and associated update tool from the manufacturer, throw it into my reverse engineering environment, tell Claude Opus 5 what my goals are, and let it churn."

One thing that became quite clear to the security engineer was that these devices lacked any decent firmware integrity protection to prevent modifying and applying a new firmware. Only the Elgato light had any defenses in this respect, and they were easy enough to circumvent.

Schlarp explains a trick with his Asus ROG Swift PG42UQ monitor to demonstrate the kind of useful utility that can be on offer with this kind of firmware modding. He found it was possible to remove an annoying pop-up warning that periodically tells the owner to run the 'pixel cleaning' process (although the engineer hasn't implemented the fix in the firmware yet). He also discovered a way to get DisplayWidget (a Windows utility) features running on his Linux system, with a shell script that can flick through certain bits of functionality like the hardware crosshair or FPS counter (which could be set up on hotkeys).

Most of what he did, though, was about proving how relatively easy it was to subvert the firmware using AI to do the heavy lifting, and, for example, disable the webcam's recording light (in the style of surveillance malware, so the user wouldn't know if the camera was secretly recording). He pulled off a similar feat with the microphone, so the mute LED could be on while the mic wasn't actually muted (this was leveraged via a 'full plaintext command shell').

Schlarp observed: "Peripherals have proven to be an ideal target for agentic RE [reverse engineering] — they're tiny computers attached to my computer, with a data connection to the host and usually a firmware update mechanism, so an agent has something to iterate against. The net outcome is better control and understanding of my machine."

Analysis: fast-tracked exploits?

(Image credit: Shure)

The key point here is how easy Claude Opus made this task. 'Owning' all five of these peripherals boiled down to 13 hours of the AI beavering away under its own steam with just shy of 100 prompts from its human overseer.

Schlarp notes that: "Hardware is almost universally 'open' for tinkering at this point with just a couple hours of mostly hands-off machine-driven labor each, and I look forward to a near future where I can add features to my webcam firmware as easily as I can to software that runs on my Linux machine itself."

However, as mentioned, there's the dark side to all this, as Schlarp makes clear: "On the other hand, as a security professional, this scares me for several reasons. I would work from the operating assumption that any device attached to a computer could have had a malicious firmware implant performed, where previously that required significant per-model investment and was stereotyped as a 'state actor' kind of activity."

In other words, the main difficulty in executing these kinds of exploits is the labor and time required, which currently limits this to individually targeted attacks on more high value targets. However, now an AI agent is capable of doing the grunt work, it makes sense that these kinds of attacks could be far more prevalent as time rolls on.

That means all those peripherals attached to your PC could be used as ways to compromise you, or your system, in the future. Schlarp informs us that he's also managed to get a root shell on a commercial Dell display, adding that: "Obviously it was never best practice to let untrusted clients touch these things, but the speed and scale at which this can be executed makes the risk so much higher now."

There's a potentially bigger threat here, too: an AI-powered worm that automatically actions this kind of reverse engineering. Schlarp explains: "It's only a tiny leap to imagine that someone could make a self-replicating piece of malware that probes its environment, relaying reconnaissance back to a smart command-and-control that actively works to push itself into accessories and IoT devices and industrial equipment found adjacent to an infected target."

There are a lot of worries about where AI could be leading us, and far more dangerous security threats looming in the future (or indeed the present) is another unfortunate prospect to say the least.

• Public confidence in AI weather forecasts is far lower than for numerical and physics-based methods
• Just 11% said they were "very confident" in machine learning weather prediction, compared with 30% for tried-and-trusted methods
• The Met Office has been using machine learning techniques in weather forecasting for some years

A survey of 6,000 UK adults has found that the majority don’t fully trust the idea of AI weather forecasts. While machine learning techniques have been a factor of weather forecasting for some years, the survey found that people are more relaxed about traditional Numerical Weather Prediction (NWP) as opposed to Machine Learning Weather Prediction (MLWP).

Predictive technology has increased considerably thanks to AI, and the Met Office has already begun to evaluate AI models for augmenting its existing forecasting methods.

However, the “confidence gap” in the results of the survey, which appears to be based around questions over accuracy, could undermine the use of AI for weather forecasting, which researchers suggest can be challenged with clear demonstrations of the methods working successfully.

AI-based weather models

The Met Office research was published in a study, Artificial Intelligence for the Earth Systems, which explores the public’s confidence in the reliability of AI-based weather forecasting and prediction. Future studies are planned to assess how attitudes change toward AI-backed weather reports.

Responses to the report are not entirely negative. While 87.7% of respondents felt confident about weather reports using NWP, the 49.4% in favour of MLWP isn’t a bad result. Rather, it demonstrates that people are comfortable with tried-and-tested methods.

Dr Edward Pope, a Met Office Science Fellow and the lead author of the paper, said the report, “deepens our understanding of current public perceptions as we approach a crucial juncture where AI-based weather models are demonstrating their potential to work alongside physics-based methods. This was a unique opportunity to compare public perceptions of established and emerging approaches to forecasting the weather."

The consequences of the “confidence gap”

The gap in confidence between the maths-based predictive forecasting and modern AI modelling is a challenge that the Met Office is addressing directly.

Dr Pope explained that “The gap in confidence and perceived accuracy highlighted in the paper demonstrates the need to clearly and transparently demonstrate the value of new approaches in ways that matter to people.”

In Pope’s view, AI practices will contribute to weather forecasts in the future, “but these improvements will only be fully realised if the public continue to have confidence, and importantly act on, the weather forecast they see.”

The Met Office research considered the “perceived accuracy” of reports, and its Chief AI Officer Professor Kirstine Dale implied that progress with AI weather modelling is subjected to evaluation and validation: “We are exploring ways of blending physics-based and AI-based modelling to deliver the forecasts that we all rely on. As with all science developments, we will robustly evaluate and validate any changes to our approach to weather forecasting before we introduce them into the model.”

Huge blazes in central and western regions of Indonesia have scorched wide areas of forests and peatland, creating a choking haze that blanketed some cities.

(Image credit: Muhammad Fajri)

Burnham, who was in Kyiv for Ukraine's Independence Day, said defense firm MBDA will release details on U.K.-made parts for the SCALP missile to Ukraine to help the country assemble the weapons.

(Image credit: Henry Nicholls)

The Trump administration is preparing to revoke the business and tourism visas of up to 200,000 foreigners who have applied for or are currently seeking asylum status in the United States.

(Image credit: Julia Demaree Nikhinson)

How are those live sports games starring the characters from The Simpsons, Toy Story and Inside Out made? With a lot of animators and tech like motion capture and face tracking.

I spoke with one of the creators behind the tool to understand exactly how it works and where it gets its sources.

When the U.S.-Iran war began, more than 1,000 dependents of U.S. Navy personnel in Bahrain were ordered to evacuate. After six months in temporary conditions, they want to know what the plan is.

(Image credit: U.S. Navy)

At D23, a panel with the stars and showrunners of Percy Jackson and the Olympians revealed a few details, including gods, rescue missions and possible deaths ahead.

Support will launch on non-folding devices first, according to the GrapheneOS Foundation.

A drug bust involving an alleged cocaine ring at Penn State has drawn attention to drug use on college campuses, but research shows the big risk to students is alcohol.

The Cupertino giant apparently will match, if not exceed, what Google and Samsung did with their phone pricing.

The Supreme Court paused a lower court's ruling that blocked key parts of Trump's order to limit voting by mail for the midterms. But U.S. Postal Service work on the order remains blocked nationwide.

(Image credit: Alex Wong)

Amid broad privacy concerns, new apps like Zuckoff help detect smart glasses in the wild.