Daily Tech Digest - February 19, 2026


Quote for the day:

“Being responsible sometimes means pissing people off.” -- Colin Powell



The new paradigm for raising up secure software engineers

CISOs were already struggling to help developers keep up with secure code principles at the speed of DevOps. Now, with AI-assisted development reshaping how code gets written and shipped, the challenge is rapidly intensifying. ... What is needed to get thrown out are traditional training methods. Consensus among security leaders is that dev training needs to be bite-sized, hands-on, and mostly embedded in developer tool chains. ... Rather than focus on preparing developers for line-by-line code review, the emphasis moves toward evaluating whether their features and functions behave securely in context of deployment conditions, says Hasan Yasar ... Developers need to recognize when AI-generated code introduces unsafe assumptions, insecure defaults, or integrations that can scale vulnerabilities across systems. And with more security enforcement built into automated engineering pipelines, developers should ideally also be trained to understand what automated gates catch, and what still requires human judgment. “Security awareness in engineering has shifted to a system-level approach rather than focusing on individual vulnerabilities,” Pinna says. ... The data from guardrails and controls being triggered can be used by the AppSec team to drive creation and delivery of more in-depth, but targeted education. When the same vulnerability or integration pattern pops up again and again, that’s a signal for focused training on a subject.


New agent framework matches human-engineered AI systems — and adds zero inference cost to deploy

In experiments on complex coding and software engineering tasks, GEA substantially outperformed existing self-improving frameworks. Perhaps most notably for enterprise decision-makers, the system autonomously evolved agents that matched or exceeded the performance of frameworks painstakingly designed by human experts. ... Unlike traditional systems where an agent only learns from its direct parent, GEA creates a shared pool of collective experience. This pool contains the evolutionary traces from all members of the parent group, including code modifications, successful solutions to tasks, and tool invocation histories. Every agent in the group gains access to this collective history, allowing them to learn from the breakthroughs and mistakes of their peers. ... The results demonstrated a massive leap in capability without increasing the number of agents used. This collaborative approach also makes the system more robust against failure. In their experiments, the researchers intentionally broke agents by manually injecting bugs into their implementations. GEA was able to repair these critical bugs in an average of 1.4 iterations, while the baseline took 5 iterations. The system effectively leverages the "healthy" members of the group to diagnose and patch the compromised ones. ... The success of GEA stems largely from its ability to consolidate improvements. The researchers tracked specific innovations invented by the agents during the evolutionary process. 


GitHub readies agents to automate repository maintenance

In order to help developers and enterprises manage the operational drag of maintaining repositories, GitHub is previewing Agentic Workflows, a new feature that uses AI to automate most routine tasks associated with repository hygiene. It won’t solve maintenance problems all by itself, though. Developers will still have to describe the automation workflows in natural language that agents can follow, storing the instructions as Markdown files in the repo created either from the terminal via the GitHub CLI or inside an editor such as Visual Studio Code. ... “Mid-sized engineering teams gain immediate productivity benefits because they struggle most with repetitive maintenance work like triage and documentation drift,” said Dion Hinchcliffe ... Patel also warned that beyond precision and signal-to-noise concerns, there is a more prosaic risk teams may underestimate at first: As agentic workflows scale across repositories and run more frequently, the underlying compute and model-inference costs can quietly compound, turning what looks like a productivity boost into a growing operational line item if left unchecked. This can become a boardroom issue for engineering heads and CIOs because they must justify return on investment, especially at a time when they are grappling with what it really means to let software agents operate inside production workflows, Patel added.


One stolen credential is all it takes to compromise everything

Identity-based compromise dominated incident response activity in 2025. Identity weaknesses played a material role in almost 90% of investigations. Initial access was driven by identity-based techniques in 65% of cases, including phishing, stolen credentials, brute force attempts, and insider activity. ... Rubin said the growing dominance of identity attacks reflects how enterprise environments have changed over the past few years, creating more opportunities for adversaries to quietly slip in through legitimate access pathways. “The increasing role of identity as the main attack vector is a result of a fundamental change in the enterprise environment,” Rubin said. “This dynamic is driven by two key factors.” He said the first driver is the rapid expansion of SaaS adoption, cloud infrastructure, and machine identities, which in many organizations now outnumber human accounts. That shift has created what he described as a “massive, unmanaged shadow estate,” where each integration represents “a new, potentially unmonitored, path into the network.” ... The time window for defenders is shrinking. The fastest 25% of intrusions reached data exfiltration in 72 minutes in 2025. The same metric was 285 minutes in 2024. A separate simulation described an AI-assisted attack that reached exfiltration in 25 minutes. Threat actors also began automating extortion operations. Unit 42 negotiators observed consistent tone and cadence in ransom communications, suggesting partial automation or AI-assisted negotiation messaging.


The emerging enterprise AI stack is missing a trust layer

This is not simply a technology problem. It is an architectural one. Today’s enterprise AI stack is built around compute, data and models, but it is missing its most critical component: a dedicated trust layer. As AI systems move from suggesting answers to taking actions, this gap is becoming the single biggest barrier to scale. ... Our ability to generate AI outputs is scaling exponentially, while our ability to understand, govern and trust those outputs remains manual, retrospective and fragmented across point solutions. ... This layer isn’t a single tool; it’s a governance plane. I often think of it as the avionics system in a modern aircraft. It doesn’t make the plane fly faster, but it continuously measures conditions and makes adjustments to keep the flight within safe parameters. Without it, you’re flying blind — especially at scale. ... Agentic systems collapse the distance between recommendation and action. When decisions are automated, there is far less tolerance for opacity or after-the-fact explanations. If an AI-driven action cannot be reconstructed, justified and owned, the risk is no longer theoretical — it is operational. This is why trust is becoming a prerequisite for autonomy. Governance models built for dashboards and quarterly reviews are not sufficient when systems act in real time. CIOs need architectures that assume scrutiny, not exception handling and that treat accountability as a design constraint rather than a policy requirement.


India Is Not a Back Office — It’s a Core Engine of Our Global Innovation

We have a very clear data and AI strategy. We are running multiple proof-of-concept initiatives across the organisation to ensure AI becomes more than just a buzzword. The key question is: how does AI create real value for Volvo Cars? It helps us become more agile and faster, whether in product development, improving internal process efficiency, or enhancing decision-making quality. India plays a crucial role here. We have a large team working on data analytics, intelligent automation, and AI, supporting these initiatives and shaping our agenda. ... It’s not just access to talent, it’s also the mindset. Indian society is highly adaptable. You often face unforeseen situations and must find solutions quickly. That agility and ability to always have a “Plan B” drive innovation, creativity, and speed. ... Data protection is a global priority. Many regions have introduced regulations, India’s Data Privacy Act, GDPR in the European Union, and similar laws in China. For global organisations, managing how data is transferred and processed across borders is a significant challenge. For example, certain data, like Chinese customer data, may need to remain within that country. Beyond regulatory compliance, cybersecurity threats are constant. Like most organisations, we experience attempted attacks on our networks. We have a robust cybersecurity team working continuously to secure both data and infrastructure.


AI likely to put a major strain on global networks—are enterprises ready?

Retrieval-heavy architecture types such as retrieval augmented generation—an AI framework that boosts large language models by first retrieving relevant, current information from external sources—create significant network traffic because data is moving across regions, object stores, and vector indexes, Kale says. “Agent-like, multi-step workflows further amplify this by triggering an additional set of retrievals and evaluations at each step,” Kale says. “All of these patterns create fast and unpredictable bursts of network traffic that today’s networks were never designed to handle. These trends will not abate, as enterprises transition from piloting AI services to running them continually.” ... In 2026, “we will see significant disruption from accelerated appetite for all things AI,” research firm Forrester noted in a late-year predictions post. “Business demands of AI systems, network connectivity, AI for IT operations, the conversational AI-powered service desk, and more are driving substantial changes that tech leaders must enable within their organizations.” ... “Inference workloads in particular create continuous, high-intensity, globally distributed traffic patterns,” Barrow says. “A single AI feature can trigger millions of additional requests per hour, and those requests are heavier—higher bandwidth, higher concurrency, and GPU-accelerated compute on the other side of the network.”


Quantum Scientists Publish Manifesto Opposing Military Use of Quantum Research

The scientists’ primary goals include: to express a unified rejection of military uses of quantum research; to open debate within the quantum community about ethical implications; to create a forum for researchers concerned about militarization; and to advocate for a public database listing all research projects at public universities funded by military or defense agencies. Quantum technologies rely on the behavior of matter and light at the smallest scales, enabling ultra-secure communication, highly sensitive sensors and powerful computing systems. According to the manifesto, these capabilities are increasingly being folded into defense strategies worldwide. ... The manifesto places these developments in the context of rising defense budgets, particularly in Europe following Russia’s invasion of Ukraine. The scientists write in the manifesto that the research and development sector is not exempt from the broader rearmament trend and that dual-use technologies — those that can serve both civilian and military ends — are increasingly prioritized in policy documents. The scientists acknowledge that quantum technologies are not inherently military tools. However, according to the manifesto, once such systems are developed, their applications may be difficult to control. The scientists argue that closer institutional ties between universities and defense agencies risk undermining academic independence. .

From pilot purgatory to productive failure: Fixing AI's broken learning loop

"Model performance can drift with data changes, user behavior, and policy updates, so a 'set it and forget it' KPI can reward the wrong thing, too late," Manos said. The penalty for CIOs, however, comes from the time lag between the misread KPI signal and the CIO's moves to correct it. Timing is everything, and "by the time a quarterly metric flags a problem, the root cause has already compounded across workflows," Manos said. ... Waiting until the end of a POC to figure out why a concept doesn't scale is clearly too late, but neither is it prudent to abandon a "trial, observation, and refine" cycle entirely, Alex Tyrrell, head of advanced technologies at Wolters Kluwer and CTO at Wolters Kluwer Health, said. Instead, Tyrrell argues for refining the interaction process itself to detect issues earlier in a safe setting, particularly in regulated, high-trust environments like healthcare. He recommends pairing each iteration with both predictive and diagnostic signals, so IT teams can intervene before the error ripples down to the customer level. ... AI pilots fail for the same non-technical reasons that have always plagued technology performance, such as a governance vacuum, organizational unreadiness, low usage rates, or "measurement theater," which is when tech performance can't be tied to a specific business value, explained Baker.


How AI agents and humans can play together in the same sandbox

Unlike traditional automation, which is rigid and rules-based, AI agents are goal-driven. They can plan, adapt, and respond to changing conditions. That makes them especially powerful for modern business processes that are dynamic by nature - processes that span systems, teams, and time zones. Another defining characteristic is endurance. AI agents don't get tired, sick, or distracted. They can operate continuously, scaling up or down as needed, and executing tasks with consistent precision. This doesn't make humans obsolete. ... Trust plays a central role here. Agents must demonstrate that they are reliable and predictable. At the same time, humans must define boundaries - what agents can do autonomously, where approvals are required, and what guardrails must always be respected. There is a fine balance to strike. Constrain agents too tightly, and you eliminate the benefits of autonomy. ... A logical approach enables AI agents to access views of data directly from source systems, in real time, without first having to replicate or move that data. For Agentic AI, this is critical: agents need live data, delivered in the shortest possible time, in order to plan, act, and adapt effectively. By abstracting physical data complexity and unifying access across sources, a logical data layer provides AI agents with fast, trusted, and governed data - exactly what autonomous systems require to operate at scale. A shared data plane provides all consumers - human or machine - with the same source of truth. It also provides context, consistency, and traceability.

No comments:

Post a Comment