Daily Tech Digest - March 03, 2026


Quote for the day:

“Appreciate the people who give you expensive things like time, loyalty and honesty.” -- Vala Afshar



Making sense of 6G: what will the ‘agentic telco’ look like?

6G will be the fundamental network for physical AI, promises Nvidia. Think of self-driving cars, robots in warehouses, or even AI-driven surgery. It’s all very futuristic; to actually deliver on these promises, a wide range of industry players will be needed, each developing the functionality of 6G. ... The ultimate goal for network operators is full automation, or “Level 5” automation. However, this seems too ambitious for now in the pre-6G era. Google refers to the twilight zone between Levels 4 and 5, with 4 assuming fully autonomous operation in certain circumstances. Currently, the obvious example of this type of automation is a partially self-driving car. As a user, you must always be ready to intervene, but ideally, the vehicle will travel without corrections. A Waymo car, which regularly drives around without a driver, is officially Level 4. ... Strikingly, most users hardly need this ongoing telco innovation. Only exceptionally extensive use of 4K streams, multiple simultaneous downloads, and/or location tracking can exceed the maximum bandwidth of most forms of 5G. Switch to 4G and in most use cases of mobile network traffic, you won’t notice the difference. You will notice a malfunction, regardless of the generation of network technology. However, the idea behind the latest 5G and future 6G networks is that these interruptions will decrease. Predictions for 6G assume a hundredfold increase in speed compared to 5G, with a similar improvement in bandwidth.


FinOps for agents: Loop limits, tool-call caps and the new unit economics of agentic SaaS

FinOps practitioners are increasingly treating AI as its own cost domain. The FinOps Foundation highlights token-based pricing, cost-per-token and cost-per-API-call tracking and anomaly detection as core practices for managing AI spend. Seat count still matters, yet I have watched two customers with the same licenses generate a 10X difference in inference and tool costs because one had standardized workflows and the other lived in exceptions. If you ship agents without a cost model, your cloud invoice quickly becomes the lesson plan ... In early pilots, teams obsess over token counts. However, for a scaled agentic SaaS running in production, we need one number that maps directly to value: Cost-per-Accepted-Outcome (CAPO). CAPO is the fully loaded cost to deliver one accepted outcome for a specific workflow. ... We calculate CAPO per workflow and per segment, then watch the distribution, not just the average. Median tells us where the product feels efficient. P95 and P99 tell us where loops, retries and tool storms are hiding. Note, failed runs belong in CAPO automatically since we treat the numerator as total fully loaded spend for that workflow (accepted + failed + abandoned + retried) and the denominator as accepted outcomes only, so every failure is “paid for” by the successes. Tagging each run with an outcome state and attributing its cost to a failure bucket allows us to track Failure Cost Share alongside CAPO and see whether the problem is acceptance rate, expensive failures or retry storms.


AI went from assistant to autonomous actor and security never caught up

The first is the agent challenge. AI systems have moved past assistants that respond to queries and into autonomous agents that execute multi-step tasks, call external tools, and make decisions without per-action human approval. This creates failure conditions that exist without any external attacker. An agent with overprivileged access and poor containment boundaries can cause damage through ordinary operation. ... The second category is the visibility challenge. Sixty-three percent of employees who used AI tools in 2025 pasted sensitive company data, including source code and customer records, into personal chatbot accounts. The average enterprise has an estimated 1,200 unofficial AI applications in use, with 86% of organizations reporting no visibility into their AI data flows. ... The third is the trust challenge. Prompt injection moved from academic research into recurring production incidents in 2025. OWASP’s 2025 LLM Top 10 list ranked prompt injection at the top. The vulnerability exists because LLMs cannot reliably separate instructions from data input. ... Wang recommended tiering agents by risk level. Agents with access to sensitive data or production systems warrant continuous adversarial testing and stronger review gates. Lower-risk agents can rely on standardized controls and periodic sampling. “The goal is to make continuous validation part of the engineering lifecycle,” she said.


A scorecard for cyber and risk culture

Cybersecurity and risk culture isn’t a vibe. It’s a set of actions, behaviors and attitudes you can point to without raising your voice. ... You can’t train people into that. You have to build an environment where that behavior makes sense, an environment based on trust and performance not one or the other ... Ownership is a design outcome. Treat it like product design. Remove friction. Clarify choices. Make it hard to do the wrong thing by accident and easy to make the best possible decision. ... If you can’t measure the behavior, you can’t claim the culture. You can claim a feeling. Feelings don’t survive audits, incidents or Board scrutiny. We’ve seen teams measure what’s easy and then call the numbers “maturity.” Training completion. Controls “done.” Zero incidents. Nice charts. Clean dashboards. Meanwhile, the real culture runs beneath the surface, making exceptions, working around friction and staying quiet when speaking up feels risky. ... One of the most dangerous culture metrics is silence dressed up as success. “Zero incidents reported” can mean you’re safe. It can also mean people don’t trust the system enough to speak up. The difference matters. The wrong interpretation is how organizations walk into breaches with a smile. Measure culture as you would safety in a factory. ... Metrics without governance create cynical employees. They see numbers. They never see action. Then they stop caring. Be careful not to make compliance ‘the culture’ as it’s what people do when no one is looking that counts.


Why encrypted backups may fail in an AI-driven ransomware era

For 20 years, I've talked up the benefits of the tech industry's best-practice 3-2-1 backup strategy. This strategy is just how it's done, and it works. Or does it? What if I told you that everything you know and everything you do to ensure quality backups is no longer viable? In fact, what if I told you that in an era of generative AI, when it comes to backups, we're all pretty much screwed? ... The easy-peasy assumption is that your data is good before it's backed up. Therefore, if something happens and you need to restore, the data you're bringing back from the backup is also good. Even without malware, AI, and bad actors, that's not always the way things turn out. Backups can get corrupted, and they might not have been written right in the first place, yada, yada, yada. But for this article, let's assume that your backup and restore process is solid, reliable, and functional. ... Even if the thieves are willing to return the data, their AI-generated vibe-coded software might be so crappy that they're unable to keep up their end of the bargain. Do you seriously think that threat actors who use vibe coding test their threat engines? ... Some truly nasty attacks specifically target immutable storage by seeking out misconfigurations. Here, they attack the management infrastructure, screwing with network data before it ever reaches the backup system. The net result is that before encryption of off-site backups begins, and before the backups even take place, the malware has suitably corrupted and infected the data. 


How Deepfakes and Injection Attacks Are Breaking Identity Verification

Unlike social media deception, these attacks can enable persistent access inside trusted environments. The downstream impact is durable: account persistence, privilege-escalation pathways, and lateral movement opportunities that start with a single false verification decision. ... One practical problem for deepfake defense is generalization: detectors that test well in controlled settings often degrade in “in-the-wild” conditions. Researchers at Purdue University evaluated deepfake detection systems using their real-world benchmark based on the Political Deepfakes Incident Database (PDID). PDID contains real incident media distributed on platforms such as X, YouTube, TikTok, and Instagram, meaning the inputs are compressed, re-encoded, and post-processed in the same ways defenders often see in production. ... It’s important to be precise: PDID measures robustness of media detection on real incident content. It does not model injection, device compromise, or full-session attacks. In real identity workflows, attackers do not choose one technique at a time; they stack them. A high-quality deepfake can be replayed. A replay can be injected. An injected stream can be automated at scale. The best media detectors still can be bypassed if the capture path is untrusted. That’s why Deepsight goes even deeper than asking “Is this video a deepfake?”


Virtual twins and AI companions target enterprise war rooms

Organisations invest millions digitising processes and implementing enterprise systems. Yet when business leaders ask questions spanning multiple domains, those systems don’t communicate effectively. Teams assemble to manually cross-reference data, spending days producing approximations rather than definitive answers. Manufacturing experts at the conference framed this as decades of incomplete digitisation. ... Addressing this requires fundamentally changing how enterprise data is structured and accessed. Rather than systems operating independently with occasional data exchanges, the approach involves projecting information from multiple sources onto unified representations that preserve relationships and context. Zimmerman used a map analogy to explain the concept. “If you take an Excel spreadsheet with location of restaurants and another Excel spreadsheet with location of flower shops, and you try to find a restaurant nearby a flower shop, that’s difficult,” he said. “If it’s on the map, it is simple because the data are correlated by nature.” ... Having unified data representations solves part of the problem. Accessing them requires interfaces that don’t force users to understand complex data structures or navigate multiple applications. The conversational AI approach – increasingly common across enterprise software – aims to let users ask questions naturally rather than construct database queries or click through application menus.



The rise of the outcome-orchestrating CIO

Delivering technology isn’t enough. Boards and business leaders want results — revenue, measurable efficiency, competitive advantage — and they’re increasingly impatient with IT organizations that can’t connect their work to those outcomes. ... Funding models change, too. Traditional IT budgets fund teams to deliver features. When the business pivots, that becomes a change request — creating friction even when it’s not an adversarial situation. “Instead, fund a value stream,” Sample says. “Then, whatever the business needs, you absorb the change and work toward shared goals. It doesn’t matter what’s on the bill because you’re all working toward the same outcome.” It’s a fundamental reframing of IT’s role. “Stop talking about shared services,” says Ijam of the Federal Reserve. “Talk about being a co-owner of value realization.” That means evolving from service provider to strategic partner — not waiting for requirements but actively shaping how technology creates business results. ... When outcome orchestration is working, the boardroom conversation changes. “CIOs are presenting business results enabled by technology — not just technology updates — and discussing where to invest next for maximum impact,” says Cox Automotive’s Johnson. “The CFO begins to see technology as an investment that generates returns, not just a cost to be managed.” ... When outcome orchestration takes hold, the impact shows up across multiple dimensions — not just in business metrics, but in how IT is perceived and how its people experience their work.


The future of banking: When AI becomes the interface

Experiences must now adapt to people—not the other way around. As generative capabilities mature, customers will increasingly expect banking interactions to be intuitive, conversational, and personalized by default, setting a much higher bar for digital experience design. ... Leadership teams must now ask harder questions. What proprietary data, intelligence, or trust signals can only our bank provide? How do we shape AI-driven payment decisions rather than merely fulfill them? And how do we ensure that when an AI decides how money moves, our institution is not just compliant, but preferred? ... AI disruption presents both significant risk and transformative opportunity for banks. To remain relevant, institutions must decide where AI should directly handle customer interactions, how seamlessly their services integrate into AI-driven ecosystems, and how their products and content are surfaced and selected by AI-led discovery and search. This requires reimagining the bank’s digital assistant across seven critical dimensions: being front and centre at the point of intent, contextual in understanding customer needs, multi-modal across voice, text, and interfaces, agentic in taking action on the customer’s behalf, revenue-generating through intelligent recommendations, open and connected to broader ecosystems, and capable of providing targeted, proactive support. 


The End of the ‘Observability Tax’: Why Enterprises are Pivoting to OpenTelemetry

For enterprises to reclaim their budget, they must first address inefficiency—the “hidden tax” of observability facing many DevOps teams. Every organization is essentially rebuilding the same pipeline from scratch, and when configurations aren’t standardized, engineers aren’t learning from each other; they’re actually repeating the same trial-and-error processes thousands of times over. This duplicated effort leads to a waste of time and resources. It often takes weeks to manually configure collectors, processors, and exporters, plus countless hours of debugging connection issues. ... If data engineers are stuck in a cycle of trial-and-error to manage their massive telemetry, then organizations are stuck drinking from a firehose instead of proactively managing their data in a targeted manner. In a world where AI demands immediate access to enormous volumes of data, this lack of flexibility becomes a fatal competitive disadvantage. If enterprises want to succeed in an AI-driven world, their data infrastructure must be able to handle the rapid velocity of data in motion without sacrificing cost-efficiency. Identifying and mitigating these hidden challenges and costs is imperative if enterprises want to turn their data into an asset rather than a liability. ... When organizations reclaim complete control of their data pipelines, they can gain a competitive edge. 

No comments:

Post a Comment