What is recurrent depth, and why are AI safety researchers worried about it?

Recurrent depth is a reasoning technique used in OpenAI's new Astra model. Instead of working through problems in a fixed, step-by-step sequence, the model can loop back through its own reasoning in less predictable ways. Safety researchers are concerned because this makes it harder to audit what the model is actually doing, which in turn makes it harder to catch errors or unsafe outputs before they cause problems.

How does the stolen session cookie attack on Claude accounts work?

When you log into a web service, your browser stores a session cookie that keeps you logged in. If an attacker steals that cookie, they can replay it to access your account without ever entering your password or your 2FA code. In the Claude case, affected accounts were personal subscriptions not managed by a company IT team, so there was no central admin who could remotely log everyone out. The attack bypassed standard login protections entirely.

What is enhanced geothermal energy, and why is it relevant to AI?

Enhanced geothermal systems generate electricity by drilling deep into hot rock and circulating water to produce steam. Unlike solar or wind, it produces power continuously regardless of weather or time of day. AI data centers run around the clock and consume enormous amounts of electricity, so continuous, low-carbon power sources like geothermal are attractive to companies like Google that have both sustainability commitments and massive energy needs.

Is the Uber and Wayve London robotaxi service fully driverless?

Not yet. A safety operator sits in the driver's seat during this initial phase of the service. The car handles driving autonomously, but a human is present to intervene if needed. Fully driverless commercial operation typically follows after a period of supervised deployment to build confidence in the system's real-world performance.

What should I do if I'm worried my driver's license was part of the ID verification breach?

Monitor your financial accounts and credit reports for unusual activity. Consider placing a credit freeze with the major credit bureaus, which prevents anyone from opening new accounts in your name. Be extra cautious about any communications asking you to verify your identity, since breached personal data is often used in follow-up phishing attempts.

AI Safety Fears, Robotaxis, and a Massive ID Breach: What's Shaping Tech Right Now

Updated on September 03, 20266 minutes read

OpenAI's latest reasoning approach is making AI safety researchers uncomfortable, a massive data breach has put over 150 million driver's license records up for sale, and London just got its first public robotaxi service. A lot happened this week, and most of it matters well beyond the headlines.

OpenAI's new reasoning model raises safety red flags

The big conversation in AI circles this week is about a technique called "recurrent depth," which OpenAI is using in its new Astra model. Unlike most reasoning models that think through problems in a linear, step-by-step way, recurrent depth lets the model loop back through its own thinking in ways that don't follow the usual sequential pattern.

Why does that matter? AI safety researchers are worried that this kind of processing is harder to interpret and audit. If a model reasons in ways that humans can't easily follow or check, spotting dangerous or wrong outputs becomes much harder. For anyone studying AI or cybersecurity, this is a live debate worth paying attention to: read about OpenAI's recurrent depth model and the safety concerns it's raised.

Google floods the zone with Gemini 3.8 Flash

Google is moving fast on its Flash model line. Just weeks after Gemini 3.7 Flash, the company released two new 3.8 Flash variants: a general-purpose model built for agentic tasks and multi-step reasoning, and a second called Flash Cyber, specifically tuned for finding and addressing software vulnerabilities.

The pace is striking. Three Flash models in six weeks suggests Google is treating rapid iteration as a competitive strategy, not just a product roadmap. For developers, the agentic version could be genuinely useful for automating multi-step workflows, while Flash Cyber is a direct play for enterprise security teams. Pricing mirrors the 3.7 Flash, at $0.75 per million input tokens. The Verge breaks down what the new Gemini 3.8 Flash actually does differently, and VentureBeat covers both variants including the cybersecurity-focused Flash Cyber.

A major ID verification service was breached, and 150M records are at risk

This is the security story of the week. Hackers appear to have broken into a significant ID card verification service, with more than 150 million driver's license photos reportedly stolen and briefly listed for sale on a dark web site before the crime site shut down. The FBI is reportedly investigating.

Here's why this one hits differently: ID verification services sit at the foundation of digital identity. Companies use them to confirm who you are when you open a bank account, sign up for a service, or take out a loan. A breach at this level doesn't just expose data; it potentially undermines the trust layer that makes online identity work. Ars Technica has a firsthand account of discovering their own license in the breach, and TechCrunch covers what's known about how the breach happened.

Stolen session cookies are bypassing two-factor authentication on Claude accounts

A separate security issue surfaced this week around Anthropic's Claude. Attackers replayed stolen browser session cookies to access paid Claude accounts, entirely bypassing both 2FA and single sign-on protections. Because the affected accounts were personal, self-serve subscriptions rather than enterprise accounts managed by IT, there was no corporate admin console that could log them out remotely.

For anyone learning about web security, this is a textbook example of why session management matters as much as authentication. Grabbing a valid session cookie lets you skip the login process entirely. VentureBeat has the technical detail on how the Claude session cookie attack worked.

London gets its first public robotaxi service

Uber and UK-based autonomous driving startup Wayve launched a commercial robotaxi service in London this week, making it the city's first. The vehicles are Ford Mustang Mach-Es, and for now, a safety operator sits in the driver's seat while the car handles driving autonomously.

Uber beat Waymo to the milestone, which is notable given Waymo's head start in US cities. The partnership between Uber (the distribution platform) and Wayve (the autonomy stack) is an interesting model: rather than each building everything themselves, they're combining strengths. For anyone interested in the business side of self-driving tech, this is worth watching. CNET covers the launch details and what passengers can expect, and The Verge has more context on how Uber beat Waymo to London.

Meta enters the real-time transcription market at a competitive price

Meta's Superintelligence Labs released Muse Voice Transcribe, a real-time speech-to-text API that handles transcription, speaker identification for more than 20 people simultaneously, and endpoint detection, all while audio is still being recorded rather than waiting for it to finish. The public API price is $0.18 per hour of processed audio.

That price point is low enough to get attention from enterprise buyers, especially for use cases like meeting transcription, call center analysis, and accessibility tooling. For developers building voice-driven apps, a competitive real-time transcription API from a major lab changes the build-vs-buy calculation. VentureBeat looks at Meta's Muse Voice Transcribe pricing and what it means for enterprises.

Palo Alto Networks acquires an AI IT automation startup for $500M

Palo Alto Networks paid around $500 million to acquire Console, an AI-powered IT service automation startup backed by Thrive Capital. The deal positions Palo Alto to expand beyond cybersecurity and into broader enterprise IT operations, and it leaves Serval, a Sequoia-backed competitor, as the main independent player in this space.

Acquisitions like this show where enterprise security spending is heading: toward AI that automates the repetitive, high-volume work of IT management, not just threat detection. If you're building skills in cloud security or enterprise software, AI-driven IT automation is an area worth understanding. TechCrunch has the sourced details on the Palo Alto Networks and Console acquisition.

Enterprises are looking beyond Nvidia for AI chips

A survey of enterprise AI infrastructure buyers found that a notably larger share plan to evaluate non-Nvidia accelerators over the next 12 months compared to those planning to evaluate Nvidia's next-generation Blackwell GPUs. Alternatives being considered include AWS Trainium, Google TPUs, AMD Instinct, and Intel Gaudi.

This doesn't mean Nvidia is losing its lead; its chips remain the default for most AI workloads. But it does show that enterprises are actively hedging, and the alternative chip ecosystem is maturing fast enough for serious evaluation. For anyone learning about AI infrastructure or cloud computing, this shift in buying behavior is worth tracking. VentureBeat details the enterprise survey results on non-Nvidia AI chip evaluations.

Google buys geothermal power for its AI data centers

Google signed a deal to purchase 400 megawatts of power from Fervo Energy, a geothermal startup, with the potential to expand to a full gigawatt. The energy would help supply power to AI data centers in Utah.

Enhanced geothermal is worth understanding as an energy story separate from solar and wind. It generates power continuously, not just when the sun shines or the wind blows, which makes it attractive for data centers that run 24/7. As AI infrastructure scales, reliable, low-carbon baseload power is something every major cloud provider needs more of. TechCrunch explains why Google's Fervo geothermal deal matters for AI data center power needs.

The common thread across this week's biggest stories is that AI is generating consequences beyond software: security vulnerabilities, energy deals, new safety debates, and physical products on city streets. The technology is moving fast, and understanding its edges, where it fails, who it benefits, and what infrastructure it demands, is becoming as important as understanding how it works.

Learn technical skills online with Code Labs Academy

Join our supportive community, unlock your potential, and embark on a rewarding career path.