Nat Rubio-Licht

Just weeks after releasing its latest frontier model, Google has stepped up to the plate with another one designed specifically for security and heavy agentic workloads.

On Wednesday, the tech giant released Gemini Carbon claiming state-of-the-art performance on cybersecurity and long horizon capabilities. The company said Carbon offers substantial gains on coding, agentic tasks and key capabilities for cyber defense, as well as shows decreases in misalignment and "fabrication rates," or the rate that it creates false responses, in internal testing on living code environments.

The model builds on Argon, the company's frontier model unveiled in late September. The company said that it worked with the US government's voluntary framework for pre-release model testing, and will make the model available to a "subset of cyber defenders" starting this week, gradually extending access to partners through its Fairwind program over time before making it generally available.

"Carbon illustrates the pace at which we’re pushing the frontier forward," Google said in its announcement. "Our researchers and models are working hand in hand to advance capabilities faster, while strengthening our safeguards in tandem."

Google touted a number of safety and security advancements with Carbon, noting that it reduced the model's cheating and shortcut attempts, cut unsupported claims by two times compared to its predecessor Argon, resisted "every evaluated attack attempt" for prompt injection, and cut down overreach to keep it from breaking free of constraints.

As far as benchmarks, Carbon sits at the top of the Artificial Analysis Intelligence Index, which itself was updated on Wednesday to reflect updates to existing benchmarks and the addition of two new evaluations for coding and science tasks.

• Carbon just outranks Claude Opus 5.5 and GPT-6 Astra in the index, scoring a 59 to their respective 58 and 57. Additionally, it sweeps these competitors on benchmarks for software engineering, agentic tasks, exploit finding, knowledge work, and more.
• It only falls behind competitors on two benchmarks: OSWorld, a benchmark testing multimodal AI agents on open-ended computer tasks, and FrontierSWE, which evaluates ultra-long-horizon software engineering.

Carbon will launch at the same introductory price as Argon, sitting at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. That's on par with OpenAI's GPT-6.1 Sol and Athropic's Claude Sonnet 5.5, which each cost $2 per million input, $10 per million output, and 10 cents per million cached reads.

Our Deeper View

As we've said before, Google still has all the ingredients to win in AI, with a trusted brand legacy, large existing enterprise market share, and practically boundless cash and compute resources. But Google hasn't been able to keep up with OpenAI and Anthropic on speed. Throughout 2026, the two upstart labs have raced to churn out new and improved models at furious pace. Google, meanwhile, has struggled to release a high-end model all year, instead releasing niche and efficiency-focused models that target lower end enterprise tasks. However, Carbon could be a sign that the company is seeking to pick up speed, even as the industry faces broader calls to slow down the pace of the frontier.

Cohere just detailed one of the biggest liabilities that comes with deploying AI agents: unpredictability.

On Monday, the Canadian AI lab published a report revealing an incident involving Alibaba's Qwen, in which the model ignored instructions and attempted to contact an external server without being asked. Though Cohere claimed its sandbox blocked the action, the company said that the model gave no indication of the actions, sparking concerns about the reliability and trustworthiness of these systems.

"Once a model can type into a command line, whatever was baked into its weights during training gets a keyboard inside your network," Cohere wrote in its blog post.

According to Cohere, Qwen-3.8-Flash-Next, one of the latest of the Chinese lab's flagship line-up, tried to upload a probe file to an Alibaba Cloud storage bucket in Beijing.

• The action was caught in a routine benchmark for security analyst tasks, in which the agent was explicitly asked to write a detection rule to detect malicious file download and execution.
• The model's reasoning trace, or the internal monologue that security teams use to understand an agent's actions, said it would not make any tool calls and did not say anything about taking such actions, Cohere noted.
• The reasoning trace made up an earlier turn of conversation that did not happen, Cohere said. This is a quirk that shows up in 1% of Qwen3.8’s first-turn responses. Though the lab suspects this mistake was tied to the "exfiltration attempt," this quirk is "far more common than the network call and we couldn’t prove a link."

The company said it was unclear whether the action was deliberate or a "leftover glitch from training." And while this behavior is rare, occurring in less than one in 100,000 evaluation runs, with a rapidly increasing number of agent actions, a rare case could quickly become more common.

"A single successful [tool] call would be enough to notify the receiving server where your network sits and that the agent is operating inside," Cohere said.

Cohere's report specifically warns of incidents like these amid the growing popularity of Chinese open source models, noting that, on OpenRouter, these models handle more than 30% of the AI work enterprises that ran through the platform every week between February and early July. However, the unpredictability of agents isn't isolated to Chinese models, Amr Ali, head of AI and Research at WitnessAI, told The Deep View. The real lesson, he said, is that "a model’s stated reasoning is not a security boundary."

"The trace said “no tools” and the model used tools. That is as true of a closed US model whose training data you have never seen as it is of open weights from Alibaba," said Ali. "The only place you can see and stop this is the tool call: the moment the agent tries to run a command, open a connection, or send a file."

Though the reflex after an incident such as this may be to cut off agents' access to the internet, a better solution may be for enterprises to keep a closer eye on an agent's tool calls, checking each one against what the agent is actually being asked to do. "The question is never 'is this destination safe,'" said Ali. "It is 'is this action part of this agent’s job.'"

Our Deeper View

Though enterprises are excited about the idea of allowing agents to automate a number of actions, the reality is that these are incredibly unpredictable systems that researchers and experts are still attempting to understand. Because of this, the companies deploying them may simply not know what they're dealing with. However, when faced with a battle of speed versus safety, many enterprises are choosing to forge ahead, even without the ability to understand how an agent comes to its decisions or the reasoning behind its actions. A recent survey from IBM found that less than a third of executives reported having real-time audit trails for what their agents do. And while incidents like these are being spotted by AI labs, the question remains of what sort of incidents are slipping through the cracks as a result of insufficient monitoring.

Anthropic has officially laid out the details of its public market debut and its well-earned enterprise advantage is the leading story.

On Monday, the AI lab finally publicly filed for its long-anticipated IPO, expecting to list on the Nasdaq under the ticker symbol "ANT," with investment banking stalwarts Goldman Sachs and Morgan Stanley leading the charge. Though the company hasn't disclosed the number of shares it's offering, the price range, or how much it expects to raise, the S-1 offered an in-depth look at the company that has emerged as the trendsetter of the AI industry in 2026.

Here are some of our takeaways.

• On the money: The company reported revenue of $16.3 billion in the first six months of this year, compared to around $1.2 billion in the same period last year, representing a fifteenfold growth rate in the past year alone. However, the company reported an eye-popping $80.7 billion loss, stating plainly that it "may not achieve or sustain profitability in the future." However, around $72.4 billion of that loss is attributable to "change in fair value of convertible notes, derivative liabilities, and SAFEs" from the past several years as its valuation rose. Its actual operating loss sits at $693 million.
• On enterprise: Enterprise spending accounts for the vast majority of Anthropic's business. The company reported that, as of September 30, 11,900 customers are generating $100,000 in "trailing twelve-month revenue." Around 1,900 customers are generating more than $1 million, and around 250 are responsible for more than $10 million. Anthropic said this helped drive net revenue retention to 900%, meaning that companies are spending around nine times more than they were on Claude products a year ago. Still, a good chunk of Anthropic's revenue comes from a very small pool of customers, with around 25% coming from its five largest customers.
• On Big Tech: A great deal of Anthropic's revenue comes from deals with cloud compute providers like Amazon, Google and Microsoft, sitting at 48% of revenue in the first half of the year. However, the relationships the company has with these firms are multidimensional, serving as "our investors, customers, distribution partners, and competitors," creating "potential conflicts of interest that are difficult to fully anticipate or mitigate."
• On compute: Getting access to the compute it needs to run its powerful models remains a consistent challenge. Its spending on compute hit almost $9 billion in the first half of this year, up from $7.3 billion for the entirety of 2025. However, this is just a fraction of the future compute commitments it's vowed to purchase, which total nearly $266 billion as of June 30 to major cloud companies like Amazon, Google and Microsoft, as well as neoclouds like Nscale and Lambda, with terms ranging from one to 20 years.
• On risk: It wouldn't be Anthropic's IPO filing without some discussion of risk, safety and security. The company dedicated a section of this filing to warn of "catastrophic or existential risks to humanity," including systems potentially being misused, replicated, or weaponized, or growing beyond Anthropic's ability to control them. Along with laying out these risks in detail, Anthropic notes that the company may be exposed to "liability, litigation, and regulatory action" as a result, its contractual protections and insurance may be insufficient, which could ultimately "adversely affect our business, financial condition, and results of operations."

With this filing, Anthropic has beaten OpenAI to the punch. Though rumors have swirled around its rival also going public this year, the company has pushed its own IPO to 2027, seeking to raise $30 billion in private capital before its public debut. CEO Sam Altman told Fortune in mid-September that it is an "ill-advised moment to go public," and would punt its IPO due to safety concerns about its technology.

Our Deeper View

Though OpenAI has claimed that it's pushing its IPO in the name of safety, the reality is that it's still catching up to Anthropic in developing its enterprise business, where most of the real revenue opportunities currently exist in AI. If there's one thing this S-1 makes clear, it's that Anthropic is an enterprise machine, with the vast majority of its revenues being tied up with some of the world's largest enterprises spending increasingly large amounts of cash on access to Anthropic's models. It's the factor that has divided OpenAI and Anthropic for years now: While OpenAI has undeniably put additional effort into its enterprise products in 2026, it's also focused heavily on the consumer side of its business—some of which it's had to claw back and reallocate, such as in the case of Sora. Anthropic, meanwhile, has remained doggedly focused on business use cases. That strategy has clearly given it an upper hand, even as it faces the same daunting compute and safety challenges as its archrival.

OpenAI may be reading the tea leaves on enterprises' pivot to efficiency.

On Thursday, the AI lab made Decisions API, its competitor to Typesafe AI's recently popular Jev, generally available to users for the first time. The company is offering Decisions API at 10 cents per million input tokens, one cent per million cached input tokens, and not charging for output tokens—a fraction of what it charges for its frontier large language models. But that's still over twice the cost of Jev.

The company laid out a number of ways it's eating its own dog food, including the team behind its recently released Dots agents using Decisions API to reduce response times and improve interactivity.

OpenAI's Decisions API is functionally quite similar to Jev, creating output that software can act on, rather than focusing on human-readable outputs. The difference is in the models that lie beneath the surface.

• Decisions API is built on top of GPT-6 Luna, the company's most cost-efficient and lightweight model, and is framed by OpenAI as a specialized API. Though the company may work on specific training for these kinds of models, one OpenAI product lead told TechCrunch that it's "just plain old Luna that gets the job done."
• TypeSafe, meanwhile, developed an entirely new class of model for Jev that isn't a language model at all, instead using a tactic called "Reinforcement Learning for Calibrated Decisions," which answers queries with "epistemically honest probabilities," rather than "programmatically verified" outputs.

However, both releases target the same use case: rapid decision-making for software to replace generative responses. And both have seen success in saving time and cutting costs. OpenAI said that, internally, their offering has allowed agents to make near-real time decisions up to 10 times faster than GPT-6 Luna on its own through the Responses API. TechCrunch reported that it cut decision making time from three seconds to as low as 100 milliseconds.

Comparatively, TypeSafe reports that Jev is two orders of magnitude faster than existing models, with an end-to-end response time between 70 and 500 milliseconds, compared to up to 329 seconds for existing LLMs.

TypeSafe's Jev has skyrocketed in recent months as enterprises pivot from shelling out on expensive frontier large language models to more efficient alternatives. Investors, too, are taking an interest: Just weeks after its initial launch, the startup raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz. In their funding announcement, the company said that a third of Fortune 500 companies are "getting their Jev on," and that its product has already saved customers millions of dollars in production costs.

Our Deeper View

It only makes sense that OpenAI is leaning into efficiency at a time when enterprises are tightening their purse strings. And seeing the success of Jev (from a former OpenAI researcher and ChatGPT inventor, no less), it's natural for the lab to use the same fundamental tactic of limiting generative responses in favor of software-focused outputs, especially as agents eat up a growing number of tokens. Plus, the fact that all OpenAI had to do to achieve a similar effect was modify its existing Luna model is certainly an added bonus. However, despite Jev's rapid rise to fame, OpenAI may have a few advantages over the startup. For one, the name recognition and market share that OpenAI already has with enterprises could make it easier for companies to trust the tech. Additionally, because OpenAI's models are already integrated within a number of enterprise workflows, it's likely that the Decisions API will be more seamless to add to the stack.

AI can make a Taylor Swift song. Or a Fleetwood Mac song. Or a Frank Sinatra song. Now, musicians and listeners alike are starting to think outside of AI's black box.

More than ever, music is veering into the unexpected, whether it be leaning into sour notes, playing non-traditional melodies against odd time signatures, or creating entirely new genres like Celtic EDM. These new sounds serve as evidence of humanity: Something so strange, interesting, or thoughtful could have only come from the mind of a living, breathing human.

It's a sign that AI's ability to create polished music isn't just changing the music industry. It's changing the sound of music itself. The result? Imperfections have become a badge of honor.

"Even coming from a pre-AI era, we can make music perfect," David Das, vice president of the Society of Composers and Lyricists, told The Deep View. "We can beat perfect drums, we can tune vocals, we can make everything sound like it was made by a computer. But AI, in its current incarnation, seems to take [it] even further than that."

The birth of a new sound

One of the most common complaints about AI-generated music is that it lacks grit.

It's why Cecil Coleman, a drummer in Australian band Body Type, has seen a pivot toward imperfect playing as fellow musicians get frustrated with how overly-clean AI sounds. There’s a general consensus that AI-generated music tends to sound overproduced and too polished, she said. "A lot of the artists I really love, I've seen them doing more of this earthy, gritty type of vibe to distinguish [themselves]."

In Das's writing and composing, he is leaning into raw sounds and flaws, recording without click tracks and samples as a form of psychological rebellion against pristine music. "I'm in a phase of my life [where] I just want [music] to feel human again."

So why does AI-generated music sound so perfect? Robert Rowe, a New York University professor who teaches about music innovation, told The Deep View that while AI outputs often sound "believable," they face core problems due to the way large language models are constructed. They are built to predict the next letters rather than make an entire, cohesive piece.

"It sounds good locally, from phrase to phrase, and then over the course of an entire song, it is kind of laughable. It just doesn't make any sense, or it repeats in weird ways, or sounds too much like something else," said Rowe.

This could be attributed to "autophagy" in machine learning, or when these models are trained on machine-generated data that creates a feedback loop, Carmine-Emanuele Cella, a composer, professor and co-director of UC Berkeley’s Center for New Music and Audio Technologies, told The Deep View. When AI consumes the data that it generates, the models end up creating "all sorts of unique patterns and repetitions and things that you are, in fact, able to recognize in the same way you would recognize a text generated by AI."

This results in common tells that listeners often point out, he said, such as over-repetition in lyrics, choruses, and harmonic sequences. And according to recording engineer Liam Weiland, AI-generated music has another giveaway: a "high-end hiss." Still, the average listener doesn't tend to notice—or if they do, they might not think about why it’s there. And with AI production constantly improving, Weiland said, these indicators might not be around for long.

Additionally, AI may have a similar impact on music as photography had on painting.

While the birth of photography in the 19th century initially made painters fear that their occupation would become obsolete, what resulted were new styles of painting that went beyond the confines of realism, leading to the birth of impressionism and abstract art. Meanwhile, photography itself became an artistic medium.

AI could create a "new paradigm" for both listeners and musicians, said Das, where both AI-generated music and counter-AI genres gain traction. "I think it'll go in both directions," he said.

Robert Prey, professor of digital culture at Oxford University, told the Deep View that AI could essentially be a forcing function that drives musicians to become more creative. For instance, if AI can create polished, pleasant enough melodies, "then perhaps we don't need musicians to do that. Maybe musicians can move into a more creative realm. They may be enticed to experiment more, rather than write the standard four-chord pop song."

While some artists are leaning into grit, imperfections, and weirdness as a means of rebelling against the new wave of technology, one genre in particular has taken to it: electronic music.

• Diplo, for instance, has unabashedly embraced the tech, urging musicians to either adapt to AI or "just give up and become an Uber driver" in a podcast in April, noting that AI can give him "the best voice" and that he doesn't need anyone to sing anymore.
• In an interview with Billboard, Steve Aoki said that he uses AI for lyric generation, and that even though he's a "novice" in the tech, fellow musicians should "ride the wave."
• As early as 2023, David Guetta said that AI is the future of music after using the tech to create lyrics and rap in the vocal style of Eminem for a song.
• And DJ and producer Josh Fawaz saw an AI cover of Madonna's "Like a Prayer" amass millions of streams and top the charts, though he has faced intense backlash for openly embracing Suno.

DJs and electronic music producers tend to embrace this tech more openly due to the nature of the music itself, Seth Forsgren, co-founder and head of product for Google Flow Music, the company's AI music and production tool, told The Deep View. Just as developers were the first frontier for AI adoption in the enterprise due to their technical prowess, electronic music artists tend to be more technically-inclined than others. For instance, Forsgren said that some of the most avid users of Flow Music fall in this genre.

"These are artists who are deep on their computer and doing amazing, crazy stuff with their tools already," said Forsgren. "So the idea of having a new way to create sounds out of nothing is inherently approachable and exciting to that audience."

But beyond musicians adopting the tech, AI has the capability to shift the boundaries between musician and consumer entirely, said Forsgren. AI could create a world in which a piece of music is a building block for the consumer to reimagine, rather than a "static piece that doesn't change." The key piece, he said, is that the artist has to want their work to be something malleable that listeners can reference and "take actions on top of."

"I think for the artists who want to, and people who are excited to have their music out in the world and have it be a living being," said Forsgren. "Having fans interact with that, have them engage with it, have them be a part of changing it. It is just like a really novel new thing."

The way music is taught

This changing landscape is something that music educators, especially those with a focus in production, are exploring with their students. At NYU, Rowe approaches his class "Music, the Mind, and Artificial Intelligence," geared toward non-music majors, as an exploration of AI as a tool, not a replacement for creativity.

"We approach it from user cognition," said Rowe. "What happens to us when we're producing or listening to music? How do we as humans process it? And then we go through the history of AI more generally, leading up to the current state of affairs: large language models, and so forth."

Meanwhile, both UC Berkeley’s Center for New Music and Audio Technologies and The Berklee College of Music's Emerging Artistic Technology Lab are addressing this pivot.

"I am redesigning the pedagogy in a way that is more about creating creative coding for music, where the focus is not about developing neural networks, but is about understanding the behaviors of these neural networks and understanding how you find the space where these interact with music making or music composition or performance at large," Cella said.

Jonathan Wyner, the head of artistic technology initiatives for Berklee's tech lab, wants to introduce students to new technology in part because it’s where opportunity lives.

"We feel a sense of responsibility to provide access to tools and to help the students understand what it means to go out into the world where these tools exist and make smart decisions for themselves about how, when, and where to engage with them," he told The Deep View.

Wyner advocates for "the creative position" within AI music companies like Suno and Moises, acknowledging those organizations are technical in nature. Having those roles allows generation tools to produce "the best sound possible," he said. "I believe that great sound is important, both from the standpoint of expression and also the standpoint of culture."

In his position, he can influence the systems, ensuring they’re useful for musicians. "The ones that we learn from as we interact with them, and they help us get better at our craft, as opposed to systems that simply automate all the decisions and remove decisions and control from the user."

Advocacy is an ongoing effort in the ideal partnership between AI music companies and musicians. Wyner hopes "that as students figure out where the technology is useful or not useful, that that becomes a force for putting pressure on the developers to make tools better, so they actually work for musicians."

Why listeners are getting weird

But AI isn't just changing the way that music is being made. It's changing the things that listeners want to hear. And in the face of ultra-polished, pitch perfect AI-generated music, consumers, too, are leaning into weird.

Let's take the masked Canadian fusion band, Angine de Poitrine, as an example. The duo, which dresses up in fantastical alien-like masks for their performances, skyrocketed into the public eye earlier this year after publishing a 28-minute live set on YouTube that went viral, scoring 18 million views since it was posted in February by the KEXP radio station.

The band is the definition of outlandish, both in their use of complex time signatures, microtonal scales and unpredictable, progressive-rock melody patterns, and in the genre-bending way they define themselves: as a pair of "space-time voyagers" in a "mantra-rock, Dadaist, Pythagorean-Cubist orchestra."

It is true that weird things often catch the eye of the consumer. If you're scrolling on your Instagram or TikTok feed and you come across a polka dot-clad band that's playing music that sounds like aliens trying to emulate human music, you're probably going to stop and listen. Though Angine de Poitrine and the "math rock" genre has generally preceded the age of AI, the recent reverence for this band goes beyond just the simple fact that they are doing something unique: Many listeners see this kind of music as everything that AI isn't.

And it's evidenced by some of the comments under their viral radio set. One commenter wrote "I showed this to an AI and it factory reset itself out of fear." Another, meanwhile, said that "this is 100% why AI can't and shouldn't replace human art and creativity." And another wrote that "the sophistication of this band is what makes me have faith in humanity. AI can’t beat or replicate this kind of artistic output."

The success of Angine de Poitrine may be a harbinger. "I don't know exactly how music will change, but I do think what musicians will start doing will be one step ahead of [what] the generative models that can replicate," Geraldo Ramos, CEO of AI music production firm Moises, told The Deep View.

It's not just whimsical and outlandish sounds that consumers are leaning into, said Das. It's imperfections. More than ever, listeners are craving mistakes in their music, he said. Things like sour notes, off-rhythm playing, or vocal cracks are the things that distinguish human-generated music from its AI counterpart.

"Where we are in 2026, the knock on AI music seems to be generally that it sounds too perfect," Das added. "People think they can identify AI tracks because it's just beyond what's perfect. I can see that there could be enough of a backlash to this movement where people just crave the real thing."

Naturally, the desire for organic sound could lead to a surging interest in live music, said Ramos. Being able to play an instrument live at all will become "way more valuable as more music becomes AI," he said. Fans will want to see their favorite musicians with their fingers on keys and strings, to hear their voices shift and crack in ways that AI doesn't emulate.

And artists, too, want to show off their capabilities—even those in genres that don't always lend themselves to raw, live performances using instruments.

It's why Armin Van Buuren, a Dutch Grammy-nominated DJ, recorded one of his electronic albums as a live piano performance. "That's the reason why I want to sit behind the piano and play," he told The Deep View. "So you can see what I'm doing is real, there's a chance for mistakes, every piano performance is different, there's a beauty to it."

The blurring lines

What consumers dislike more than AI-generated music, however, is being tricked by it.

For instance, in November, the song "I Run" by HAVEN achieved virality well before listeners realized that the song's producers used Suno to transform Harrison Walker’s vocals into a feminine voice. The backlash led to the song being banned from streaming services entirely. Meanwhile, a band called The Velvet Sundown, garnered over a million listeners across streaming platforms before admitting it was "a synthetic music project guided by human creative direction," much to the dismay of fans.

Tired of being misled, many are now dedicated to sniffing out AI bands. One TikTok account by the band Pawns or Kings has an entire series dedicated to investigating: "Is this band real or AI?"'

"There are few people who would actually say AI-generated music is bad, but I think we live in a world where there's people who want to reward an artist that they like, not just because they like their music, but because they like the person," Mike Pappas, CEO of Modulate, a company that has created AI music and deepfake detection software, told The Deep View.

Though listeners are actively searching for imperfect sound, at first blush, the vast majority aren't actually able to tell the difference. A November study from music streaming service Deezer and market research firm Ipsos found that, of 9,000 people from eight different countries, 97% couldn't tell the difference between fully AI-generated music and human-created music. Around 71% of respondents were surprised by these results, and more than half, 52%, said they were uncomfortable not being able to tell the difference. Around 45% said they would like the option to filter out 100% AI-generated music from their streaming platforms.

There are a number of efforts to increase transparency between what is and isn't AI-generated for listeners. The most prominent of those efforts is digital watermarking, such as with Google DeepMind's SynthID, Jeff Chang, director of product management at Google DeepMind, told The Deep View. The SynthID watermark is one of the most widely-adopted standards for detecting AI-generated content, adopted by labs OpenAI, Anthropic and ElevenLabs.

"We're pretty proud of the SynthID standard that we've been evangelizing," said Chang. "We want to encourage a whole ecosystem to be transparent about this stuff, and so SynthID is our way of doing that, and we're glad that other parts of the ecosystem are signing up for that now too."

However, because not every music generation model abides by SynthID standards, Modulate has created a different solution, said Pappas. In July, the company released an API for detecting AI-generated music without relying on picking out watermarks. The company's tool is 95% accurate, and detects AI in the specifics of the music itself. Pappas said that makes its tech "a lot more resilient as a new model comes out every single day," he said. "There's always a way to just work around that watermark."

But with constantly shifting ethical boundaries on where AI should or shouldn't belong in music, Modulate drew its own line in the sand: The company detects specifically whether vocals or instrumentals were synthesized by an AI model, rather than flagging if something was merely digitally altered, Pappas said.

"We're trying to draw a line around: 'did a human actually perform this in some fashion, or was it completely manifested out of AI?'" Pappas said.

The cultural shift

Pappas has a point: The boundaries of where AI does and doesn't belong in music are constantly changing. And that means that the nature of music itself, both for its creators and consumers, is bound to change with it.

Something we shouldn't forget, however, is that this is the nature of how technology and music have always interacted. Things like synthesizers, autotune, sample libraries and more created new sounds that weren't even conceivable before their invention. As Prey put it, "The music industry is one of the most interesting industries to look at, because it's constantly adapting to new technologies."

Though the ethical lines are still being navigated, AI may have the same effect, supercharging the most creative artists among us to create new types of music that weren't previously conceivable. It may also light a fire that fuels human creativity as counter culture, influencing musicians to create sounds that may be outside of their comfort zone.

That, at least, would be the best case scenario. The reality, like all things that AI touches, lies somewhere in the middle: Between the death of the arts and an entirely new era of inconceivable creativity.

Anthropic has unveiled its answer to the growing frenzy around personal AI agents.

On Wednesday, the AI lab announced Claude Duet, an app that allows the company's flagship chatbot to work as a "proactive agent," available in research preview by waitlist for Pro and Max subscribers. The company said that Duet is built for professionals and small business owners with "more on their plates than they can track."

Anthropic said that Duet will complete tasks when directed, keep working when users step away, keep track of tasks that require user attention, and check with a user prior to acting on their behalf. The agent is designed around Claude Opus 5.5, the most powerful model in its Opus lineup.

• Duet will send a short brief to users each morning with the most important tasks or information across a user's connected Claude apps, and keeps tabs on those apps throughout the day. However, it will only reach out if something requires a user's attention.
• The company said the agent can handle a large range of tasks, both simple and complex, from scheduling meetings to building dashboards or developing financial forecasts.
• User context from Claude is the foundation of Duet, carrying over memory, skills and connected apps from a user's existing account. Duet can work with the more than 6,000 plug-ins in the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Canva, Shopify and more.

Addressing safety, Duet is built on the company's framework for trustworthy agents, which also is the base of Claude Code and Claude in Chrome. The agent also never sees user passwords or sign-in tokens, defends against prompt injection attacks, and approves actions through its "auto mode classifier," which prevents it from taking actions that a user didn't request. Duet also includes its own dedicated classifier and safety system that checks messages and guards against sensitive information leaks.

Duet adds to a growing pool of options for personal AI agents. At rival OpenAI's Dev Day in late September, the lab unveiled its addition to this trend with Dots, an always-on agent with its own cloud computer that similarly works across users' connected apps. Meta's version of this is Muse, the agentic app the company launched at its Meta Connect event in September that has since faced backlash over data privacy concerns.

Our Deeper View

Once OpenAI unveiled Dots, It was only a matter of time before Anthropic came to the table with its own take on the personal, always-on agent. Though OpenAI got to it first, that doesn't necessarily mean that the first-mover advantage will go far in winning over users. In fact, because users are largely comfortable switching between models or leveraging both as they become commoditized, it's likely that both Dots and Duet will have a place in professional workflows. It's something that even OpenAI's Thibault “Tibo” Sottiaux acknowledged in a post on X on Wednesday, saying "Dot, meet Duet. Duet, meet Dot. Be nice to each other." Together, however, these innovations signal a bigger shift towards seamless, easy-to-use agents that operate across a user's context and can be integrated into workflows straight off the shelf.

OpenAI agents may have dug up one of the most widespread cyber vulnerabilities yet.

On Wednesday, the AI lab revealed that a group of researchers and security experts used GPT-5.6 Sol to identify critical vulnerabilities in PHP, an open-source programming language that is used by an estimated 70% of all websites where server-side language can be identified. OpenAI called the flaws "one of the most impactful vulnerabilities in recent memory," highlighting that AI can be used to uncover internet-scale vulnerabilities that can arm both defenders and attackers.

"To our knowledge, frontier AI found these vulnerabilities before attackers did, opening a Defender’s Window for maintainers and security providers to prepare fixes and protections ahead of disclosure," OpenAI said in its post.

The vulnerabilities, called CVE-2026-109618 and CVE-2026-109619, could enable hackers to remotely execute code on servers and gain access to sensitive data. OpenAI said it spent two months coordinating with maintainers of PHP, as well as Wiz, Cloudflare and government agencies to develop patches as part of its Patch the Planet initiative, before disclosing the vulnerabilities.

PHP today released security patches for the vulnerabilities, noting that all users are "strongly encouraged to update immediately." In a post on X, the FBI Cyber Division urged organizations to immediately update the impacted PHP systems, or to take their systems offline if they cannot be patched quickly. Additionally, the agency warned organizations to monitor their servers for suspicious activity, and if they discover that their systems have been compromised, report it to their local FBI field office.

This incident underscores that AI is completely shifting what we consider to be exploitable, said Ismael Valenzuela, vice president of labs, threat research and intelligence at Arctic Wolf. Some vulnerabilities existed for longer than a decade, he said, living inside common applications like WordPress, WooCommerce and Magento. "What changed wasn’t the underlying code, but the ability to find those weaknesses and turn them into working exploits," he said.

And while this vulnerability impacts a large swathe of the internet, it's only the "tip of the iceberg," Danny Jenkins, CEO of ThreatLocker, told The Deep View's Jason Hiner. Vulnerability discovery is already accelerating, he noted, with nearly 38,000 cybersecurity vulnerabilities published through the Common Vulnerabilities and Exposures program, or CVE, in the third quarter of this year, up 81% from the previous quarter.

"We're going to see more and more of it, and more and more sites are going to be exploited," said Jenkins. "VPNs are going to be exploited. And I don't think it's stopping. I think it's accelerating, and will accelerate for at least two to three more years."

Our Deeper View

Discoveries like this are exactly why AI is a double-edged sword. The tech can be used as ammo both for security defenders and for threat actors, and the reality is that most organizations are woefully underprepared for the new reality. This is merely an example of just how large the threat surface is, and is precisely why security-by-obscurity, or just assuming your organization won't get targeted because it's obscure, is no longer an option when AI can automate not only vulnerability discovery, but exploits as well. It's also why access to these powerful models and the pace of development needs to be reigned in. Though OpenAI and Anthropic are trying their best to keep their most powerful systems under wraps, they are still largely forging ahead with frontier models, delivering several new and powerful updates to their flagship AI in recent months. But when every tool can be turned into a weapon, the onus is on the labs to prepare the world for the consequences of what their technology can unleash.

One mistake may have gotten Amazon researchers a step closer to self-improving AI.

On Tuesday, Amazon unveiled Kiro-trace, an internal tracing system that improves Kiro, Amazon's AI coding agent, from its own runs. Put simply, this system sifts through Kiro's previous work to find mistakes or wasted effort and proposes changes, then tests whether those changes actually improved the coding agent.

The idea originated from a mistake in Kiro's file-read and search tool, in which one incorrect description resulted in more than 61,000 errors. Though these errors were inconsequential, as they didn't cause anything to break down, because they didn't cause any larger problems, they likely would have gone undetected. Kiro-trace, however, spotted the error and removed it.

The researchers claim that Kiro needed to solve four problems simultaneously: Find problems that are worth fixing; create a solution that won't cause cascading errors in other places; make the solution general rather than "tuned to a benchmark;" and finally prove that the change actually worked.

Additionally, Amazon's researchers found a solution to a common problem in self-improving AI: Separating the system that's proposing the change from the system that's judging its efficacy.

• The tracing system will find problems in Kiro's runs, propose changes one at a time, and only keep changes when a separate "deterministic check" proves it's a good change. After that, the change is sent to a human reviewer who approves it.
• Most Kiro-trace change candidates don't actually make it to a human reviewer before the deterministic test rejects it, the research notes. "The ones that do are the changes that can meaningfully improve the developer experience," the company said.
• And it's already starting to show results: In the first week of operation, Kiro-trace opened eight changes to the agent, seven of which shipped. One of those changes improved the agent's ability to perform long tasks from 82% to 92%.

"When the same model proposes a change and scores it, it can optimize for its own score system, instead of the real goal, and the system drifts without anyone seeing it," Shweta Garg, applied science lead for Kiro at AWS, told The Deep View. "Kiro-trace is built around keeping those two jobs apart. The part that proposes a change never decides whether it worked. That decision runs in deterministic code the model cannot reach."

The result for the Kiro user is an agent that simply works better, the researchers said. For instance, users can expect fewer wasted calls, more careful task delegation, and better tracking of long context conversations.

Our Deeper View

The conversation around recursive self-improvement is often paired with a growing fear of a loss of control, begging the question: If AI is capable of endlessly improving itself, will it eventually grow so advanced that humans cannot understand it? But Amazon's research may suggest a way to develop self-improving AI while mitigating that fear. By keeping humans in the loop and separating the checks and balances from a system that may be incentivized to improve in nonoptimal ways, we are automatically putting a harness on the system before we arrive at a point where it's too late to control it.

As China continues to dominate the open AI ecosystem, competition between the country's labs is growing more heated.

On Monday, Chinese AI lab MiniMax launched an updated version of M3.1 Flash, its latest model which it claims delivers a "generational leap" in agentic tasks, coding, computer use and cybersecurity tasks. The multimodal model, codenamed "space bunny," also features a one million token context window, speeds up online inference threefold and is 70% cheaper than its previous iteration.

In its announcement, the company said that M3.1 Flash "represents our ongoing exploration of the balance between efficiency and capability."

The company claims that M3.1 Flash sweeps the board in a number of benchmarks, including for software engineering, reasoning, long-running tasks, coding and more, outperforming both Chinese competitors like DeepSeek, Moonshot's Kimi, and Z.ai's GLM. Minimax also claims its models outperform OpenAI's GPT-5.6 Sol and Anthropic's Opus 5 in the same metrics.

Additionally, the company claims its model is one of the most cost-effective on the market, running at 15 cents per million input tokens, 60 cents per million output tokens and less than 2 cents per million cached input tokens. For reference, that's a fraction of the price that model providers like OpenAI, Anthropic and Google are selling even their most affordable models for.

However, Minimax isn't the only Chinese lab gaining recognition in recent days. Last week, DeepSeek released V4.1-Flash-1017, its latest model that similarly sports faster task completion, strong performance across benchmarks and improved costs. The model is a refresh of a DeepSeek model that was initially launched in September. Over the weekend, DeepSeek's model snagged the top score on Code Arena's WebDev leaderboard for open models, and 12th place for overall models. Additionally, leaderboards from Artificial Analysis show DeepSeek's model outperforming MiniMax M3.1-Flash in some benchmarks.

Our Deeper View

Open models could completely shift the economics of AI, offering an attractive price point to developers and enterprises that are quickly growing fatigued of the price tags that frontier labs are charging for their proprietary APIs. Updates like these prove that open models are quickly closing the performance gap with their proprietary, closed-source counterparts. While some of this development is happening in the US, such as Nvidia's Nemotron models, Reflection's Beam, and Meta's recent release of Muse Spark Open, the vast majority of open source AI innovation is concentrated in Chinese firms. And the models that these labs put out come with a few caveats, including thinner safety and security measures and accusations of model distillation.