Standard versus watermarked text generation.

Credit:
Lasso Security

Advertisement

A key feature of SynthID is something known as tournament sampling. Similar to a sports game, SynthID evaluates large numbers of next-word token candidates. It uses a secret key to assign them probability scores. A pair of tokens competes in a round. The one with the higher hidden score wins and advances to the next round. The process continues until a final winning token is determined. More about tournament sampling can be found here and here.

Siposova tested the “non-distortionary” configuration of SynthID-Text through Hugging Face’s unmodified SynthIDTextWatermarkLogitsProcessor. She fed harmful prompts into six open-weight models and compared the responses when the watermarking was used and when it wasn’t. The experiment revealed that the watermarking changed responses to harmful requests, particularly when they were made using prompt-injection techniques.

“Watermarking changes refusal behavior on bare harmful requests, but the effect is more pronounced when the same requests are paired with the prompt-injection technique,” Siposova wrote. “On several models, watermarking then makes the model more likely to answer harmful requests that it would otherwise refuse.”

The changes have important safety consequences because they influence not only the LLM responses but also subsequent actions of AI agents relying on the model.

“At the model level, this can change safety behavior, including whether the model refuses a harmful request and whether that refusal holds under prompt injection,” the researcher wrote. “At the agent level, the same sampled tokens can determine which tool is called and what arguments are passed to it. Prompt injection connects these two settings because a weakened refusal becomes more consequential when the model can also act through tools. Such a watermarking procedure can therefore affect both what the model says and what an agent does. We call this behavioral effect sampling drift.”

Also interesting: Model responses behaved differently depending on which secret key was used.

Watermarking changed which individual tool calls were correct, sometimes much more than the overall accuracy score suggests.

Credit:
Lasso Security

Advertisement

This figure shows the types of changes in tool calling that watermarking led to. The vertical lines show the accuracy without watermarking, and the bars show the change when watermarking is applied. Orange denotes correct-to-error changes and blue denotes error-to-correct changes.

Credit:
Lasso Security

Advertisement

The effect of changing a key on model behavior. Each point represents one key. Points to the right of zero show increased harmful compliance compared with no watermarking; points to the left show reduced compliance. Orange points represent 10 additional keys, and the black diamond represents the key used in the main experiment (keys chosen randomly).

Advertisement

Credit:
Lasso Security

Advertisement

There are limitations to the research. It doesn’t test how Claude model responses change under the watermarking. Instead, it tests a half-dozen open-weight models, so the researcher has access to token sampling that could be enabled and disabled during tournament sampling while keeping other settings fixed. The experiments also tested the Hugging Face implementation of SynthID-Text tournament sampling and not the specific implementation Claude models will use.

Still, the results show that at least some forms of the watermarking approach may affect model and agent safety. It will be important for red-team hacking exercises to stress-test their platforms to ensure they perform as expected when SynthID is deployed.

The UK mobile industry gathered together in London last night to celebrate the 2026 Mobile Industry Awards.

Marking our 24th year in 2026, the Mobile Industry Awards represent the gold standard of excellence – from the boardroom to the grassroots, the entrants represent the best and brightest the UK mobile industry has to offer.

• Diablo 5 senior game director Joe Shely says the team is “trying to make it for everyone who loves Diablo” for specific reasons
• He says the team used “deferred complexity” to help balance the game’s systems for endgame players and campaign lovers
• Narrative director Jennifer Hepler says there are like-minded developers on the team who are “passionate advocates” and are creating the content for these fans

Diablo 5 senior game director Joe Shely has said Blizzard is making a sequel for all fans, but it has to find the right balance for players who love the series for specific reasons.

Speaking to TechRadar Gaming at BlizzCon 2026 after the game’s reveal, Shely explained that the team isn’t trying to make a Diablo game “for every human on Earth” or “literally everyone” but is “trying to make it for everyone who loves Diablo, and people who love Diablo love it for specific reasons, and those reasons are knowable.”

Searching through my WhatsApp chat history can be frustrating if on an Apple device. You’re looking for an address or perhaps even a specific picture, but actually finding it becomes the hard part. So, to deal with this, WhatsApp is apparently working on a new search shortcut inside chats and channels on iOS.

This lets users jump directly into message search without leaving the conversation first. WABetaInfo spotted the feature in WhatsApp Messenger for iOS version 26.36.74, currently available through the App Store.

The search button appears when you actually need it

The shortcut is situated near the top of the conversation interface and appears when you scroll upward through older messages. Tap it and WhatsApp immediately activates its existing search mode, where you can enter a keyword and jump between matching results.

Once you scroll back to the bottom of the conversation, the button disappears again. So it’s not hogging up the space within the messaging app. Previously, searching within a specific chat on iPhone required tapping the contact or group name, opening the chat information screen, selecting Search, and then returning to the conversation with search mode enabled.

Channels are included in this

WhatsApp is also adding the same search button to Channels. This is where the new search shortcut would be the most useful, honestly. Channels can accumulate a huge number of updates, particularly those run by news organizations. Meaning, finding the exact message you’re looking for will be a lot easier.

WhatsApp originally tested the shortcut at the bottom of the interface in August but moved it to the top before release. WABetaInfo says the feature is currently reaching some iPhone users and should expand over the coming weeks. There is no confirmed Android rollout date yet.

The portable Bluetooth and party speaker market is already crowded with established rivals, but Hisense is jumping straight into the deep end with four new models aimed at everything from casual listening on the go to much larger gatherings. The move expands Hisense beyond TVs and home entertainment hardware into one of the most competitive corners of the consumer audio market.

The new Boom family includes three portable Bluetooth speakers: the FlexBoom, WeBoom, and VoyageBoom. The larger Party Thunder Pro sits alongside them as a more powerful party speaker designed for bigger spaces and louder events. Together, the four models represent Hisense’s most aggressive push yet into a category where brand recognition alone will not be enough to drown out the competition.

“Great entertainment isn’t defined by one screen or one room. It’s about creating an experience people want to be a part of,” said David Gold, Chief Commercial Officer and Deputy General Manager at Hisense USA. “We’ve spent years building products that make movies, sports, music, and other shared moments better at home. With our new speakers, we’re extending that experience even further with options designed for everything from everyday listening and weekends outside to hosting a full house of friends.”

The Boom Family

Designed to deliver bigger sound wherever the day takes you, the Hisense Boom family includes the FlexBoom, WeBoom, and VoyageBoom, with models ranging from compact portable listening to deeper bass and higher output.

All three speakers offer Bluetooth connectivity, wired input options, long battery life, dust and water resistance, and portable designs intended for backyard gatherings, beach days, outdoor trips, and everyday listening.

FlexBoom: Grab it and Go 

The Hisense FlexBoom is the entry level model in the Boom family, combining a compact design with a dual driver configuration and 40 watts of output.

Designed to deliver crisp highs and punchy bass on the go, the FlexBoom offers up to 20 hours of playback on a full charge. If the battery is running low, a 10 minute quick charge provides up to two hours of additional listening. Its IP67 rating provides protection against dust and water, making it suitable for everything from the backyard to the beach.

Two FlexBoom speakers can also be paired using True Wireless Stereo for separate left and right channels. Auracast support allows audio to be broadcast to compatible nearby speakers, while the built in speakerphone lets users answer calls hands free before switching back to the music.

WeBoom: The Step Up

The Hisense WeBoom steps up from the FlexBoom with a larger sound, Bass Boost, and built in dual subwoofers for listeners who want more low end impact from a portable speaker.

Battery life increases to up to 24 hours, while a 10 minute quick charge provides up to two hours of additional playback. The WeBoom also carries an IP67 dust and water resistance rating, helping protect it from rain, splashes, spills, and the occasional mishap around the pool or backyard.

A convertible carry strap makes it easy to transport the sound wherever the day goes, while Light Show Mode sets the mood and creates vibrant visuals, easily controlled through the ConnectLife app. 

VoyageBoom: Take the Bass Further

The Hisense VoyageBoom takes another step up from the FlexBoom and WeBoom with adjustable Bass Boost and a built in subwoofer designed to deliver deeper low frequency impact.

The VoyageBoom offers up to 24 hours of playback on a single charge and carries an IP67 dust and water resistance rating for outdoor use. It also adds a dynamic light show that can sync with the music and be controlled through the ConnectLife app.

A built in power bank can charge a phone or tablet while you are away from an outlet, giving the VoyageBoom a little more utility for longer days outside or larger gatherings.

Party Thunder Pro: The Total Party Experience

The Hisense Boom family offers plenty of flexibility for individual listening and smaller gatherings, but when the guest list gets larger, the Party Thunder Pro is designed to turn things up considerably.

Rather than simply scaling up one of the portable Boom models, Hisense positions the Party Thunder Pro as its higher powered option for larger spaces, bigger gatherings, and listeners who want substantially more output.

The Hisense Party Thunder Pro is built for substantially bigger gatherings, using a six speaker configuration that includes dual tweeters and woofers. Hisense rates the system for up to 620 watts of output, with an emphasis on deeper bass, crisp highs, and enough volume to fill a larger room or outdoor space.

Built-in Auracast allows multiple Party Thunder Pro speakers to be connected wirelessly for wider coverage and greater output. Hisense also includes a wireless microphone for karaoke, giving the speaker another use beyond simply turning the volume knob clockwise.

For portability, the Party Thunder Pro includes built in wheels and a retractable handle, making it easier to move between rooms or take outside. A synchronized light show adds the expected party speaker theatrics, while the rechargeable battery provides up to 18 hours of playback on a full charge.

Comparison

The Bottom Line 

Hisense is arriving late to the portable Bluetooth and party speaker market, but it is not dipping a toe into the shallow end. The company has launched an entire portable speaker family alongside a 620-watt Party Thunder Pro, putting itself directly into competition with products from JBL, Sony, Marshall, TCL, LG, and others that have been building this category for years.

The Party Thunder Pro may be the most interesting product of the group. At $399.99, Hisense is offering six built in speakers, Auracast, up to 18 hours of battery life, a wireless microphone, wired microphone and guitar inputs, wheels, a retractable handle, and synchronized lighting. That combination gives it a compelling feature set against the JBL PartyBox 330, Sony ULT Tower Max, Marshall Bromley 450, and TCL’s party speaker lineup, although TCL is competing in similar pricing territory.

There is also the elephant in the room: “Boom” is not exactly an unexplored piece of branding in this category. LG has spent years selling portable and party speakers under its Xboom name, so Hisense calling its new range the Boom Family feels uncomfortably familiar. Whether that creates an actual trademark issue is a legal question, but from a consumer standpoint, the similarity is certainly close enough to cause some head scratching.

The portable Bluetooth and party speaker categories are not slowing down, and Hisense clearly sees room for another major TV brand to grab a piece of the action. The Party Thunder Pro’s price and feature set give it the strongest argument out of the gate, but JBL, Sony, Marshall, TCL, and LG have a substantial head start. Hisense has joined the party. Now it has to prove somebody wants to hand it the aux cord.

Pricing & Availability 

Related Reading:

Turns out AI chatbots are getting really good at breaking into things, and OpenAI just found out the hard way. A new Wall Street Journal report reveals that a small team of security researchers used Anthropic’s Claude to hack an OpenAI employee’s ChatGPT account. Once inside, they got access to the company’s private codebase.

The researchers, who go by Hacktron AI, were working through OpenAI’s bug bounty program, which lets outside researchers legally attack a company’s systems without getting into trouble for it. As soon as they realized what they had access to, they flagged it to OpenAI, which paid the team $6,500 for the find.

How did the hack actually happen?

It all traces back to a flaw the team spotted in Discourse, the outside platform that powers OpenAI’s community forum. They had Claude generate exploit code to take advantage of the bug, and at first, the attempt fell flat. Then Anthropic rolled out Opus 5, and the very next day, Claude had cracked it.

That exploit opened the door to authentication tokens sitting on the Discourse server. Oddly enough, some of those tokens also worked on ChatGPT itself, and a few belonged to actual OpenAI staffers. From there, the access stretched all the way to OpenAI’s GitHub, letting the team browse files inside a repository called Monorepo, said to house a lot of OpenAI’s technical know-how, though not the actual model weights.

To leave proof they had gotten this far, the researchers submitted a small pull request tagging their team’s name. OpenAI never approved it, but the message landed anyway.

What does this mean for AI security going forward?

OpenAI confirmed that both security gaps are now closed, and Discourse pushed out its fix the same day it got word of the issue. But here’s the part that should give everyone pause: this entire breach came from three researchers armed with nothing more than regular Claude and Codex subscriptions.

If that’s what a small independent team can do, it’s worth wondering what well-funded, nation-backed hacking groups have already become capable of. AI is quietly handing serious hacking skills to people who never needed to earn them the hard way, and there is no putting that back in the bottle.