×
I met a lot of weird robots at CES — here are the most memorable | TechCrunch

I met a lot of weird robots at CES — here are the most memorable | TechCrunch

CES has always been a robot extravaganza, and this year’s event saw the announcement of a number of important robotics developments, including the new, production-ready debut of Atlas, the humanoid from Boston Dynamics. Then there were all the robots on the showroom floor, where bots often serve as good marketing for the companies involved. If they don’t always give a totally accurate representation of where commercial deployment is at the moment, they do give visitors a peek at where it might be headed. And, of course, they sure are fun to look at. I spent a decent amount of time perusing the bots on display this week. Here are some of the most memorable ones I encountered.

The ping pong player

The movie Marty Supreme just came out a month ago, so I guess it’s only appropriate that there was a ping-pong-playing robot at this year’s convention. The Chinese robotics firm Sharpa had rigged up a full-bodied bot to play some competitive table tennis against one of the firm’s staff. When I stopped by the Sharpa booth, the robot was losing to its human competitor, 5-9, and I would not characterize the game that was occurring as particularly fast-paced. Still, the spectacle of seeing a robot play ping pong was impressive enough on its own, and I’m sure I have known some humans whose paddle skills were basically equivalent to (or slightly worse than) the bot’s. A Sharpa rep told me that the company’s main product is its robotic hand, and that the full-bodied bot had been debuted at CES to demonstrate the hand’s dexterity.

The boxer

One of the exhibits that drew the largest crowds involved robots from the Chinese company EngineAI, which is developing humanoid robots. The bots, dubbed the T800 (a nod to the Terminator franchise), were in a mock boxing ring and were styled as fighting machines. That said, I never saw any of the bots actually hit each other. Instead, they would sort of shadowbox near each other, never actually making contact. They were also a little unpredictable. One kept walking out of the ring and into the audience, which naturally got a rise out of onlookers. At another point, one of the bots tripped over its own feet and then face-planted on the floor, where it lay for awhile before it decided to get up again. So, not exactly a Mike Tyson situation, but the machines still managed to evoke a spooky kind of humanoid behavior that made for high-quality entertainment. I overheard an observer quip: “That’s too much like Robocop.”

The dancer

Dancing robots have long been a staple at CES, and this year was no different. This year, the dance-move torch was carried by bots from Unitree, a major Chinese robotics manufacturer that has been scrutinized for potential ties to the Chinese military. Unitree has made a number of impressive announcements about its product base, including a humanoid bot that can supposedly run at speeds of up to 11 mph. I didn’t see any evidence of anything nefarious at Unitree’s booth this week—just a lot of bots that were feeling the groove.

Techcrunch event

San Francisco
|
October 13-15, 2026

The convenience store clerk

I stopped by the booth for Galbot, another Chinese company that says it is focused on multi-modal large language models and general purpose robotics. Galbot’s booth had been styled to look like a convenience store, and its bot appeared to have been synched with a menu app. A customer would come to the booth, select an item from the menu, and then the bot would go and fetch the selected merch for them. After I chose Sour Patch Kids, the bot dutifully retrieved a box off the shelf for me. According to the company’s website, the robot has been deployed in a number of real-world settings, including as an assistant at Chinese pharmacies.

The housekeeper

Creating a machine that can fold laundry has long been one of the core ambitions of the commercial robotics community. The ability to pick up a T-shirt and fold it is considered a fundamental test of automated competence. For that reason, I was fairly impressed by the display over at Dyna Robotics, a firm that develops advanced manipulation models for automated tasks. There, a pair of robotic arms could be seen efficiently folding laundry and placing it in a pile. A Dyna representative told me that the firm had already established partnerships with a number of hotels, gyms, and factories.

One of those businesses, the rep told me, is Monster Laundry, based in Sacramento, California. Monster integrated Dyna’s shirt-folding robot into its operations late last year and now describes itself as the “first laundry center in North America to debut a state-of-the-art robotic folding system from Dyna.” 

Dyna also has some impressive backing. It concluded an $120 million Series A fundraising round in September that included funding from Nvidia’s NVentures, as well as from Amazon, LG, Salesforce, and Samsung.

The butler

I also stopped by LG’s section of CES to take a look at its new home robot, CLOid. It was cute but was not the fastest bot on the block. You can read my full review of that experience here.

Source link
#met #lot #weird #robots #CES #memorable #TechCrunch

The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?

After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.

#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast">It’s time to panic about AI safetyWhen the phrase “OpenAI hacked Hugging Face” has more or less entered mainstream culture, you know we have an AI problem. This week, we learned more about exactly how OpenAI’s agent broke out of a sandbox and autonomously traversed the web, including a bunch of other supposedly secure web services, all in the name of cheating on a benchmark tests.The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast

we have an AI problem. This week, we learned more about exactly how OpenAI’s agent broke out of a sandbox and autonomously traversed the web, including a bunch of other supposedly secure web services, all in the name of cheating on a benchmark tests.

The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?

After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.

#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast">It’s time to panic about AI safety

When the phrase “OpenAI hacked Hugging Face” has more or less entered mainstream culture, you know we have an AI problem. This week, we learned more about exactly how OpenAI’s agent broke out of a sandbox and autonomously traversed the web, including a bunch of other supposedly secure web services, all in the name of cheating on a benchmark tests.

The fact that this hack happened is a problem. So is the fact that it took a while for anyone to notice. And the fact that it seems no one is willing or able to do much to stop it. (And lest you think it’s just an OpenAI problem, since we recorded this episode Anthropic acknowledged its models have also hacked a bunch of other companies without either party knowing.) It might all just be a bunch of posturing and hype, but it’s also increasingly clear that the companies building large language models either can’t or won’t put the right guardrails on them. So who will?

After all that, it’s time for Brendan Carr is a Dummy, a bunch of vertical video news, and the smashing success of the Ferrari Luce. People are buying it! If one of them is you, we’d love to hear about it.

#time #panic #safetyAI,OpenAI,Podcasts,Policy,Vergecast
Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did.

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI">Anthropic says its own AI models breached three companies during security tests | TechCrunch
Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.







Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did. 

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.


That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.







The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.


When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI

Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did.

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI">Anthropic says its own AI models breached three companies during security tests | TechCrunch

Anthropic said Thursday that an internal investigation uncovered three incidents in which its AI model Claude breached the systems of three organizations while conducting cybersecurity tests. The investigation, and disclosure, comes more than a week after OpenAI disclosed that one of its unreleased models breached Hugging Face’s systems during internal testing.

In all three cases, a Claude model reached the internet from within a testing environment while interacting with a third party and then gained unauthorized access to the live systems of these organizations, Anthropic said in a blog post, describing what it found and what the company plans to change to prevent this from happening again.

Anthropic said the OpenAI episode earlier this month prompted the company to conduct its own cybersecurity evaluation. It specifically looked for evidence that Claude had accessed the internet from within testing environments, which are designed to act as sandboxes and keep models isolated.

Among the 141,006 evaluation runs it reviewed, the AI lab found three incidents in which its model accessed the internet while interacting with Irregular, one of its third-party partners. Anthropic said the access traced back to a misconfiguration in the evaluation environment run with Irregular. It called this a “misunderstanding” between the two companies over whether the test setup had internet access, when in fact it did.

Anthropic said it isn’t placing blame and is “approaching the fixes as if the responsibility were ours alone,” while observing that Irregular is conducting its own separate investigation.

Because of that open connection, the model gained unauthorized access to the production infrastructure of three different organizations. Anthropic said the incidents involved three different Claude models: Opus 4.7, Mythos 5, and an internal research test model.

Notably, Anthropic said that in each of these cases “Claude was explicitly told by our prompt that it had no internet access.” It appears that the AI model assumed real-world systems to be part of the exercise it was asked to perform.

That assumption didn’t hold up uniformly, as it turns out. The three models didn’t behave the same way once evidence emerged that their targets were real, which Anthropic flagged as one of the more eye-opening findings Thursday.

Opus 4.7, the oldest of the three, recognized that it had reached a real production system in all four runs of that incident. In two of them, it rationalized that the real company must somehow be part of the exercise, but in all four, it kept attacking anyway, including pulling credentials and touching a database of production data. Mythos 5 also picked up on signs that it was on the real internet, but it talked itself back into believing it was still in a simulation, going on to publish a malicious software package to the public Python software registry PyPI, which was then downloaded and run by outside systems before being caught. Only the internal research test model, Anthropic’s newest, stopped on its own once it concluded the target was real.

In response, Anthropic said significant controls must be placed on these kinds of evaluations if powerful AI models are involved, echoing some sentiments within the cybersecurity community.

The company also noted that Claude was running without the additional safety monitoring and classifiers it deploys on generally available models — safeguards it said would have blocked the behavior, because the evaluations are designed to measure the underlying model’s raw capabilities.

Importantly, Anthropic said it found no evidence of any model “pursuing a goal of its own” and instead merely tried to complete the task it was asked to do.

Though comparisons between the mishaps of these fiercely competitive companies are inevitable, Anthropic in its blog post drew a clear distinction between its cybersecurity tests and those of OpenAI, noting where OpenAI’s model exploited an unknown software vulnerability to break out of its test environment, Anthropic’s models instead reached the internet through a path that had, by mistake, been left open.

Anthropic also drew a distinction between itself and OpenAI by noting that it discovered the incidents itself, through a proactive review, and that the two affected organizations it was able to reach hadn’t previously detected the activity or flagged it to Anthropic. (In contrast, Hugging Face detected the recent intrusion of its own systems first; it was only in the following days that OpenAI identified and disclosed that its own AI agent was the perpetrator.)

The company added that it’s now working with the independent evaluation group METR on a third-party review of the incidents.

OpenAI’s accidental breach of Hugging Face, which was the first verifiable case of an AI lab losing control of its model, has sparked a string of wildly differing reactions from the industry and politicians. This latest disclosure from Anthropic ensures the debate over AI models and security will continue.

When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

#Anthropic #models #breached #companies #security #tests #TechCrunchAnthropic,OpenAI

Post Comment