For years, Synthesia’s pitch was to use AI to help enterprises create interactive training videos, in minutes, for a fraction of the cost normally associated with developing corporate educational materials. The startup’s newest product is making a different bet — that the real moat in enterprise AI isn’t just generating content, but in proving that it worked. 

On Wednesday, the British startup launched Roleplay Sessions, an interactive training product where employees can practice high-stakes conversations like sales pitches, performance reviews, and customer complaints with an AI avatar that talks back, pushes back, and then scores them against a rubric. 

Roleplay is the first release under a broader “Sessions” platform Synthesia plans to expand into other formats, including job interviews and candidate screening, TechCrunch has exclusively learned. 

The launch comes as enterprises take a harder look at whether their AI spend is paying off. Citing a meta-analysis of learning research, Synthesia argues that most corporate training — including, implicitly, its own core video product — stops short of actually changing behavior. The trainings tend to inform and demonstrate, but what really pushes someone into actually learning something is doing that thing, along with feedback.

“Video performs way better than text or sending out a document,” Synthesia CEO and co-founder Victor Riparbelli told TechCrunch. “But for most things, we learn the best by actually practicing something rather than just reading it.”

Roleplay also moves Synthesia into a more defensible position. The company’s avatar and voice technology are proprietary, but the underlying reasoning intelligence is OpenAI’s. By tacking on a layer that includes rubrics, performance data, and analytics, Synthesia is positioning itself less as an AI avatar company and more as a performance-management platform with a video front end. 

“From a managerial executive layer, we’re seeing huge interest in mapping out the talent in their company in a way that’s been really difficult to do before,” Riparbelli said. “If you have your entire sales team practice with an AI role player, you can get very granular data on how your sales force is doing overall.”

Synthesia already has a few early Roleplay customers with scaled commercial rollouts, including “one of the top three companies by market cap in Europe, one of the top five Fortune 100, and one of the biggest recruitment companies in the world,” per a company spokesperson. 

The two most popular use cases involve training sales teams and leadership — programs like practicing how to sell a product or have a difficult conversation. Much of it is soft skills training, Riparbelli said. 

As with Synthesia’s main product, customers can either create their own training programs on the platform or enlist the help of Synthesia’s consultants to build a bespoke solution that relies in part on whatever training documents and context the company already has. 

Roleplay is an enterprise offering for now, but the goal is to offer it more widely to small businesses, prosumers, and potentially educational institutions in the next few months as inference costs continue to fall. Riparbelli believes products like Roleplay represent the next phase of enterprise AI adoption, where companies care less about whether AI looks good in a social media demo and more about proving it generates measurable business outcomes. 

It’s been a big year for AI acquisitions — so big that most of them barely register anymore. Anthropic and OpenAI have each gone on buying sprees, snapping up developer tooling, AI services shops, and product-testing startups to convert model capability into enterprise revenue and extend their reach faster than the other. Which is what made a weekend rumor about Anthropic acquiring robotics startup Physical Intelligence stand out. It spread exceedingly fast, even after a denial from Physical Intelligence’s CEO

Part of that ties to who’s involved. Physical Intelligence isn’t some obscure robotics shop. It was co-founded by Lachy Groom, an investor-operator whose star has been on the rise in Silicon Valley in recent years; it has raised more than $1 billion (and was reportedly in talks this spring for another $1 billion round at an $11 billion valuation); and its π0.5 model is apparently among of the more widely used robot brains in robotics research.

As it turns out, the rumor wasn’t completely spurious. Anthropic and Physical Intelligence actually did hold acquisition talks this spring, according to The Information, so tech blogger Robert Scoble — whose weekend post on X set off the frenzy — may have gotten the specifics wrong without being wrong that something had happened.

Physical Intelligence’s response to the rumor mill wasn’t the world’s most vigorous denial, it should be noted. According to The Information, Physical Intelligence CEO Karol Hausman told employees the reports weren’t true via a Slack message containing a gif of a character from “The Office” shaking her head no.

Groom, for his part, did not respond to TechCrunch’s request for comment, sent Monday night.

Anthropic has made four known acquisitions this year; OpenAI has been more aggressive, acquiring at least 17 companies since 2023. Both are also, of course, now preparing to go public. Anthropic confidentially filed for an IPO on June 1, followed by OpenAI a week later, setting up what could be two of the largest U.S. stock debuts in history.

So why robotics, why now? The likeliest answer is that physical-world understanding may be a prerequisite for superintelligent systems, and no amount of internet text can substitute for it.

OpenAI’s own history here is instructive. It built an early robotic hand that could solve a Rubik’s Cube, then shut the entire robotics group down in 2021, with co-founder Wojciech Zaremba later saying the approach was missing pieces needed for real superintelligence. The team came back in 2024, quietly building a humanoid robotics lab in San Francisco, before CEO Sam Altman made it official in late May, announcing “OpenAI Robotics” was hiring and describing a near-term focus on robots for infrastructure work, with a personal robot for everyone as the long-term goal.

Anthropic hasn’t built anything resembling OpenAI’s hardware lab. What it has done is publish a string of research pieces through its internal group that stress-tests frontier capabilities for safety purposes. That included Project Fetch last November, where Anthropic staff tested how much Claude could help non-experts program a robot dog and a second phase in June that, according to Anthropic, found a newer model completed the same tasks roughly 20 times faster than the best human-plus-Claude team from the year before.

Buying an existing team with robotics expertise would let Anthropic skip years of work. There’s a possible complication, though. Physical Intelligence was founded in San Francisco roughly two years ago by Groom, former Google researchers, and professors from Stanford and Berkeley, and its early investor base looks a lot like OpenAI’s own, including Khosla Ventures and Thrive Capital. Founders Fund — also a major OpenAI investor — was reportedly involved in Physical Intelligence’s newest funding round earlier this year.

In fact, OpenAI is itself an investor in Physical Intelligence, so it isn’t just a peripheral player; it’s a stakeholder in a company that its chief rival was reportedly in talks to buy very recently.

That raises questions around whether OpenAI’s early investment came with any information rights, or a right of first refusal over a sale to a competitor — the kind of protective provisions that large strategic investors sometimes negotiate for precisely this scenario.

That leaves open the possibility that if Physical Intelligence is actually in play, OpenAI — already a shareholder, already close to Groom, already trying to ensure it bests Anthropic in robotics — may have the more obvious claim to it than Anthropic does. We asked OpenAI these questions earlier today and the company didn’t respond.

OpenAI admitted Tuesday that one of its AI models breached the systems of Hugging Face, the unaffiliated AI hosting platform, during an internal cybersecurity test that went awry. The models reportedly escaped their isolated testing environment and reached Hugging Face’s systems from there. Hugging Face initially attributed the breach to an “external AI agent.”

In a blog post published Tuesday afternoon, OpenAI detailed the steps that led the models to compromise the service.

“After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark⁠ of cyber capabilities,” the post reads.

In particular, the breach appears to have focused on ExploitGym, a publicly hosted benchmark measuring models’ ability to execute attacks based on existing vulnerabilities. Benchmarks like ExploitGym are commonly used in model training to refine specific skills, but this is the first known incident in which that testing resulted in an actual cyberattack.

In this case, the model in question should not have even had internet access, outside of a specific tool that enabled models to install software packages they might need to complete their task. Instead, the model was able to find an undisclosed vulnerability in the package-installer program, which it used to access the broader internet at will.

“The models were hyperfocused on finding a solution for ExploitGym, going to extreme lengths to achieve a rather narrow testing goal,” OpenAI’s post reads. “After gaining Internet access, the models inferred that Hugging Face potentially hosted models, datasets and solutions for ExploitGym. Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation.”

Ultimately, the models found vulnerabilities in Hugging Face’s infrastructure that allowed them to “obtain test solutions directly from Hugging Face’s production database,” effectively providing the answers to the benchmark.

For Hugging Face, the apparent result was a sophisticated and aggressive cyberattack, with “many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services,” as the company stated in its initial disclosure.

OpenAI has identified and reported the vulnerabilities in the package installer and is working with Hugging Face to investigate the incident further. The company also said it would implement new controls on both model testing and the related infrastructure, meant to prevent similar incidents in the future.

It’s unclear whether OpenAI will face any legal consequences as a result of the breach, although it’s likely that the models’ actions violated the Computer Fraud and Abuse Act.

Nevertheless, the result is an unusually vivid illustration of the power and dangers of frontier AI models operating on long time horizons. As OpenAI researcher Micah Carroll posted in response to the news, “If this doesn’t convince you that misalignment risks are going to be a key concern going forward, I don’t know what will.”