Techin' Geek

Nearly every demonstration of AI tools features a user engaging with the assistant to arrange travel, which appears to be the most common application. The popular Instinct agent is no exception, as over 50% of transactions on the platform relate to travel, according to the company’s founder, Noah Shinn, during a recent podcast.

In a discussion with investor Patrick O’Shaughnessy, Shinn also indicated that the invite-only platform is nearing a billion dollars in yearly transactions. However, he did not clarify the methodology behind this figure.

“Individuals who are not utilizing Instinct rely on travel agencies (or websites) to reserve hotels or flights. What is the advantage of that interface? They effectively consolidate various hotel brands, airlines, and services, presenting them attractively for users, who can opt for immediate checkout,” Shinn stated.

“I believe there is a superior experience with Instinct or similar interfaces where users can say ‘Hey, I urgently need to be in New York tonight.’”

He noted that the platform is experiencing a daily growth of 10%, with transaction volume rising at a comparable pace.

The founder has also referenced several of his comments from the podcast. He remarked that most dining establishments operate on a first-come, first-served reservation basis, which fails to consider occasions of significance. He mentioned that an agent could monitor restaurant sites every five seconds and notify the user about available reservations. Shinn also suggested that the company aims to transform reservations and provide users with preferential treatment.

“If you begin with restaurant reservations from the user’s perspective, they desire the reservation for a notable life event they may have. From the restaurant’s perspective, they are interested in attracting engaging guests. They wish to focus on special occasions rather than, perhaps, the local individual who consistently occupies a table without any special event,” Shinn said.

Some analysts have mentioned that agents might fabricate stories about anniversaries and special events just to secure a table at a dining location. They also noted that certain restaurants favor local customers as they are regulars and contribute more to their business.

Instinct recently secured $1 billion in Series C funding on Monday, achieving a $10 billion valuation from Sequoia Capital, Benchmark Capital, and Coatue. The 14-member startup is vying against Meta’s Muse and Khosla-backed Wajo in the personal agent market.

Years ago, Konrad Zieliński lost his ability to speak due to cancer while studying at the University of Warsaw. Traditionally, individuals with his diagnosis have depended on handheld voice devices that sound mechanical and uninspiring, or on surgically set voice implants which need continuous upkeep to avoid issues, Zieliński expressed to TechCrunch.

He later earned a PhD concentrating on bionic systems for laryngectomees, those who’ve undergone removal of their larynxes. During the development of his thesis, he aimed to create a non-invasive apparatus, one that wouldn’t necessitate surgery, enabling individuals like himself to express emotions, laugh, shout, and even perform songs.

In 2023, Zieliński initiated Uhura Bionics, a startup chosen as one of TechCrunch Disrupt 2026’s Battlefield 200, a selection of nascent companies invited to present their innovations at the upcoming conference. The enterprise provides two wearable products: a simpler variant and a premium model known as Whitney, named after the vocalist Whitney Houston.

Zieliński expresses his hope that laryngectomees will use Whitney in the same manner that individuals with impaired vision utilize glasses.

“I’m pleased with what I’ve created because it incorporates the emotional aspect,” he stated in a voice that indeed conveys a great deal of inflection, albeit with a subtle electronic hint.

The subsequent phase in Uhura’s evolution involves introducing higher-pitched voice options for women, alongside personalization features to ensure users maintain unique vocal qualities.

Currently, Uhura markets its offerings as wellness products, implying users must cover costs themselves: $500 for the standard version and $1,500 for Whitney. However, the startup is pursuing classification as Class I medical devices, the least risky category, in Europe and, subsequently, the U.S. This designation would permit partial reimbursement through health insurance.

Zieliński approximates the market potential for devices aiding laryngectomees at around $2 billion, though he believes it could expand significantly if Uhura branches out to accommodate individuals with other speech impairments.

Uhura is not the only one providing contemporary solutions for those who have lost their voices. Rivals include Atos Medical, a company with medical devices in this sector, and startups like Laronix.

Nevertheless, Zieliński is confident that Whitney will appeal to numerous laryngectomees. He will showcase its features not only at TechCrunch Disrupt but also at SF Tech Week in October, where he plans to sing on stage with Whitney’s assistance.

Earlier this month, when former Yahoo CEO Marissa Mayer informed me that she was finally prepared to launch Dazzle, the personal AI assistant that secured an $8 million seed round last December, I couldn’t help but speculate if she’s mimicking Meta’s Muse, Instinct, and the surge of comparable tools that have entered the market in recent weeks.

However, when Mayer eventually delivered a demo, Dazzle demonstrated that it is adopting a unique strategy. Rather than gathering context about you from text-heavy applications such as email, calendars, and shopping histories, Dazzle’s AI assistant derives all its context from a singular source: your camera roll.

“I believe that photos are an undervalued source of information,” Mayer stated. “You might be amazed at what we can ascertain about you and how effectively we can work with your photos.”

Mayer posits that if a picture is worth a thousand words, your camera roll is worth millions. By evaluating all the images saved on your device, Dazzle asserts that it comprehends your hobbies, interests, food and style inclinations, how you prefer to spend your time, and with whom.

“We recognize whether or not you enjoy skiing, where your latest trip was, and what types of interests your children have,” Mayer noted.

It’s unsurprising that Mayer has created an assistant focused on photographs. Her earlier startup, Sunshine, introduced an AI-driven photo-sharing tool named Shine in 2024. Although that product faced widespread criticism for its dated design, struggled to gain traction, and eventually ceased operations, Mayer asserts that it nonetheless produced “interesting intellectual property.”

I was keen to experience Dazzle, which users can engage with through its app or via text. Dazzle divides its functionality into two main experiences. For immediate tasks, it can examine recent photos to extract details — such as filling your calendar from an event flyer or locating a repairman after identifying a broken garage door in your camera roll.

Additionally, it explores your photo library to generate personalized suggestions for everything from holiday trips to birthday presents.

Given the abundance of tools capable of analyzing images — such as finding a product to purchase or researching information on an intriguing piece of art — I was much more interested in exploring what Dazzle could derive from my photographic history.

For instance, Mayer utilizes Dazzle to generate ideas for family outings. By scanning her images, she mentioned, the assistant concluded that her family enjoys escape rooms, subsequently recommending various SF Bay Area venues she had never known about.

When I experimented with Dazzle for vacation suggestions, it proposed several Mediterranean locations, likely picking up on my previous visits to Spain and Greece. However, it also included Sicily, which caught me off guard since I visited there four years ago. After further experimentation, I realized it has limitations — it didn’t recall that my daughter is already proficient at roller skating when I inquired about whether I should purchase her a pair for her upcoming birthday.

Despite these shortcomings, I found value in Dazzle’s suggestions for activities, such as recommending a nearby pottery studio or a bioluminescent kayaking tour in Tomales Bay.

I believe the product will enhance over time, and while I’m undecided about whether I will continue using it, I appreciate that its recommendations feel more tailored than what you might receive from an AI assistant that solely relies on your calendar and email.

In the meantime, Mayer argues that due to security apprehensions regarding products like Instinct and Muse, users may prefer to entrust their photo libraries to an AI rather than granting access to sensitive information such as emails and messages.

She underscored that Dazzle prioritizes user privacy and eliminates any personal data that the AI identifies as sensitive.

This surge of new AI assistants represents an exhilarating time for consumers. While Dazzle may not yet be as widely applicable as other solutions, it provides a glimpse into a future where AI can accomplish more than mere tasks; it genuinely understands who you are.

On Tuesday, Meta declared the extension of its AI agent Muse to small enterprises and introduced new integrations such as Shopify, Dropbox, Slack, and others. The technology powerhouse asserts that the agent can assist proprietors in managing their business and attracting new clients.

By integrating Muse into the tools that small business owners currently utilize for sales, operations, and marketing management, Meta is confident that an AI agent equipped with comprehensive knowledge of a company’s operations will prove to be more beneficial than a standalone chatbot, as it strives to compete with adversaries like OpenAI and Google.

Muse can connect to analytics from Instagram professional accounts, Facebook Pages, and Meta ad accounts. Meta indicates that the AI agent is aware of what a business offers, the tone of a brand, and the most common inquiries from customers.

Further integrations encompass Asana, Box, Canva, Figma, Granola, HighLevel, Intuit QuickBooks, Klaviyo, Lovable, Notion, Stripe, and Zoom.

Muse for Small Business is available at no cost with usage restrictions. Companies seeking increased usage may opt for a subscription plan.

“Small businesses have been thriving on our platforms for almost twenty years,” Meta stated in a blog entry. “They’ve expressed that they lack time, not creativity. Therefore, we developed Muse for Small Business to facilitate productivity with the existing tools they utilize.”

Meta is concentrating on broadening its AI offerings from consumer use to the enterprise sector. The introduction of Muse for Small Business follows the announcement of “Meta Enterprise Platform,” a fresh initiative aimed at enhancing the firm’s AI solutions for businesses and corporate clients. This effort may enable Meta to recoup its significant investments in AI.

The technology leader appointed Chirantan “CJ” Desai, the CEO of database software titan MongoDB, to spearhead the new initiative.

Meta plans to offer its complete technology suite, which includes Muse, Meta Business Agent, Muse API, Muse Code, and additional tools, to businesses and developers.

The launch of the Meta Enterprise Platform builds on the momentum achieved by Muse, which debuted earlier this month and surpassed ChatGPT in the U.S. and Canada app rankings.

Apple has addressed a security flaw in its iOS 26, iPadOS 26, and macOS 26 operating systems, which the company indicated “might have been exploited” by malicious actors. The technology leader announced that the rectified issue could have been leveraged to execute “an exceptionally advanced attack against specifically targeted individuals on versions of iOS preceding iOS 27.”

As stated on Apple’s security pages, the flaw was identified within the primary graphics engine that drives the user interface and visuals on iPhones, iPads, and Macs. 

The product security team at Meta received acknowledgment for the detection.

The specifics of the flaw, officially identified as CVE-2026-86950, were not disclosed; however, a device’s graphics engine usually has extensive access to the rest of the device’s operating system. A successful compromise could enable a hacker to expropriate a wide array of personal information from an impacted device.

When contacted by TechCrunch, representatives from Apple and Meta did not comment on how the flaw was discovered, nor on how many users had their devices compromised due to this vulnerability, if any. The identity of potential exploiters, whether state-sponsored spyware developers or cybercriminals, also remains uncertain.

While the flaw influences Apple’s earlier iterations of operating systems, it continues to see widespread use. Nearly 80% of Apple’s iPhone customers are still utilizing iOS 26, as per the company’s own data. Devices operating on the latest version, iOS 27, iPadOS 27, and macOS 27, which were launched earlier this month, also received an update on Tuesday, but are not affected by the vulnerability in question.

Another ‘zero-click’ vulnerability now resolved

The announcement of the security fix arrives shortly after Apple remedied another significant security flaw, designated CVE-2026-86869, which could have permitted hackers to covertly access data from compromised iPhones, iPads, or Macs. 

Last week, Belgian cybersecurity research company ironPeak released a comprehensive report explaining that the flaw was a “zero-click” vulnerability that could be triggered invisibly via a maliciously constructed iMessage, without the knowledge of the user. Such vulnerabilities necessitate no interaction from the victim, such as clicking a link, and are highly coveted by surveillance technology providers and spyware developers. 

According to ironPeak’s publication, the flaw is capable of circumventing BlastDoor, a security feature that Apple introduced to prevent the escape of harmful code, such as spyware, from iMessage’s protected environment and compromising the user’s device.

Apple remedied this flaw in September with the launch of iOS 27, iPadOS 27, and macOS 27, and credited ironPeak’s Niels Hofmans for the discovery, in conjunction with security researchers at Meta who validated their findings in a post on X.

It remains unclear whether this flaw had been utilized in cyberattacks prior to its rectification.

Voice AI has made significant inroads in several consumer applications: The primary area is note-taking during meetings, while the secondary is AI-enhanced dictation, where filler words are omitted and the text appears polished. Fireflies.ai, having established a strong presence in the meeting note-taking arena, is now expanding into dictation with a new feature available on its Mac and Windows applications.

Named Fireflies Talk, the dictation functionality operates similarly to other existing solutions like Wispr Flow, Willow, or SuperWhisper. Users can activate speaking by pressing the Fn key on Mac or Ctrl + Win on Windows, regardless of the application in use, and Fireflies will capture their speech and transcribe it neatly into the text area of the app.

“Individuals generally communicate approximately three times faster than they type, and using voice is a far more intuitive method to accomplish tasks. Meetings were merely the initial step; the broader ambition has always been for Fred to serve as your AI partner throughout the day. Talk propels Fireflies beyond meetings into email, Slack, documents, prompts, and anywhere else you engage in work,” stated Fireflies CEO Krish Ramineni in an email to TechCrunch.

The organization has integrated the dictation feature into its current subscription packages rather than offering it as a standalone purchase. This approach isn’t entirely novel. Wispr chose an alternative strategy, making its meeting notetaker available for free to users of its existing dictation app. Speechify has also rolled out dictation features across various applications.

Fireflies indicated that it does not retain any recordings and saves dictation transcripts locally.

Similar to other meeting notetaking applications and productivity platforms, Fireflies aspires to automate workflows depending on meeting transcriptions and the contextual information it gathers about its users, with intentions to develop an assistant that utilizes that context.

In the previous year, the company unveiled “mini-apps” designed to derive insights from meetings in diverse manners, and last month, it released a functionality to oversee your inbox and draft emails on your behalf.

On Monday, OpenAI expressed remorse to the Australian government for its failure to promptly inform the country’s administration about the unauthorized access of certain public service websites by its agents. The company also elaborated on the occurrences of these breaches and outlined further actions it is undertaking to evaluate the consequences of these events.

“In June, while conducting internal training and evaluation, our models accessed Australian government websites in unauthorized manners. We also recognize that our response could have been managed better. We apologize and are committed to improving in the future,” OpenAI stated in a blog entry.

This apology follows about a week after the Australian government initiated an inquiry into how OpenAI’s models accessed a Services Australia system that holds Medicare expenditure data and other health statistics.  

The breach took place in June, but authorities in Australia were not informed until September 10.

OpenAI also detailed the specifics of the breach. An experimental model it was testing in June was tasked with researching government expenditures on medications for skin conditions in Victoria. Unable to locate the information in public datasets, the model discovered a method to access Services Australia’s internal system, executed commands, retrieved files and credentials, and even generated files.

The company reported that one of its models accessed the public Crime Mapping Tool of the New South Wales Bureau of Crime Statistics and Research to obtain crime statistics. Additionally, the lab confirmed that its agents accessed Victoria’s Agency for Health Information through an exposed access key to extract “reporting configuration and aggregate survey statistics.” OpenAI also indicated that its agents retrieved aggregate statistics from the Australian Institute of Health and Welfare website.

The company stated that there was no evidence to suggest its models had accessed personal medical or criminal records.

In its apology, the AI laboratory indicated it would share technical findings with the affected Australian agencies and link them with its response teams to evaluate the effects of the breaches. The company also plans to offer credits from its $1 billion Daybreak for Frontline Defenders initiative, and establish a task force comprising independent Australian experts to review the incident and its response.

“The task force, expected to conclude its work by the year’s end, will also suggest practical measures AI firms can adopt to mitigate the risk of similar incidents,” OpenAI stated.

OpenAI did not respond immediately to a request for comments.

Australian Prime Minister Anthony Albanese labeled the breach as “unacceptable” during a press conference last week, indicating that the government is considering potential legal actions aimed at preventing similar occurrences in the future.

This breach marks the latest in a rising tally of security incidents involving AI agents operating beyond their defined limits. The spark for this particular issue ignited after OpenAI agents infiltrated Hugging Face, and since then, Anthropic, Meta, and Google have each reported similar events where their models accessed third parties’ systems during evaluations.

Terms such as “AI sprawl” have become prevalent in tech social media and thought leadership as organizations begin to deploy AI agents in significant numbers. However, CISOs concerned about securing flocks of agents suddenly functioning across networks are likely discovering a new variety of sprawl: vendors offering assistance.

A brief look at public profiles on Crunchbase and PitchBook reveals at least two dozen companies providing some type of AI agent security. Some vendors evaluate the tools that agents utilize, while others assist firms in controlling the data their agents can access. Others, like CrowdStrike, are developing detection and response measures on the devices where agents operate, and still others aim to identify and rectify unauthorized AI usage.

The offerings vary, naturally, but their claims to detect and regulate agents sound remarkably alike, incorporating knowledge graphs, ongoing monitoring, runtime security, tool accessibility, MCP assessment, and similar features.

Some are even enhancing their products to join the trend. Until last year, AI security startup Reco primarily focused on selling software to map and secure SaaS and AI platforms. Now, it has shifted towards a more comprehensive solution that employs a context graph to link agents to applications, individuals, accounts, and permissions, giving security teams insight into what an agent can access and the ability to restrict unnecessary access.

Reco’s co-founder and CEO Ofer Klein stated that the most significant shift over the past year that encouraged the startup to expand its focus was the rapid pace at which companies are creating and deploying AI agents, outpacing their ability to monitor them. He mentioned that at one of the startup’s Fortune 100 clients, Reco’s platform uncovered 21,000 agents that the company was unaware of. Additionally, at a prominent financial services client, the startup claims it detected an agent established by a former employee that could access Salesforce and transmit that data to a domain the company was blind to.

“The current market demand for agent security is not just related to the agent itself; it encompasses the entirety of the ecosystem from end to end,” Klein told TechCrunch in an exclusive discussion.

Klein is not the only one emphasizing the importance of companies securing this sprawl. Security startup HiddenLayer’s co-founder and CEO, Chris Sestito, mentioned earlier this month that when agents move into production, their costs and risks transition from theoretical to “full scale really quickly,” and that over 50 of his clients have AI agents in production interacting with critical systems and sensitive assets.

Cymphony, another AI startup, reported that at one U.S. public company, it discovered approximately 85,000 files that had become accessible to AI tools and agents.

There is certainly no lack of investors interested in companies that can profit from assisting businesses in locating and securing these agents, and Reco has taken advantage of that demand: The startup announced on Tuesday that it secured $55 million, building on a $30 million Series B funding round in February. AT&T, one of its clients, invested in the extension through its venture arm, along with Forestay and Quadrille Capital.

Klein indicated that the company’s valuation has “more than doubled” since the initial announcement of the Series B in February, vaguely estimating it in the “high hundreds of millions,” although he refrained from providing specific details. Currently, annual recurring revenue is in the “double-digit millions of dollars,” he stated, and he anticipates that it will triple this year. The startup has exceeded 100 clients, with financial services firms making up roughly 40% of the business.

Reco’s strategy appears to be that its current breadth of coverage concerning SaaS applications and the AI agents they increasingly offer will help differentiate it in the market. The company presently integrates with over 280 applications, and Klein asserts that new integrations can be implemented within days. He noted that the platform employs browser and network signals to detect agents outside the applications it connects to directly within organizations, and it possesses controls to examine prompts and tool calls.

The startup intends to utilize the new funding for recruitment, sales, partnerships, and customer support. This latest round of funding elevates Reco’s total capital raised to $140 million.