The AI stories enterprise IT leaders actually read.
Volume 342, Issue 4 · 20 articles from 6 issues
A CISO op-ed lists four evidence-based questions boards now expect CISOs to answer about AI risk.
In this SC Media Perspectives column, TrustCloud CEO Sravish Sridhar argues boards now expect CISOs to answer four questions about AI with evidence rather than assurances: whether humans stay in the loop on consequential decisions given new AI-enabled attack paths; whether compliance with the EU AI Act, ISO 42001, and NIST AI RMF is continuously demonstrable rather than reconstructed annually; how AI risk from internal use and vendor supply chains is governed under one program; and whether AI has measurably lowered risk and cost.
It cites Verizon's 2026 DBIR (45% of employees now use AI on corporate devices) and McKinsey data showing only about 37% of firms see measurable EBIT impact from AI.
AI's real constraint isn't intelligence but the inability to measure the economic value it creates versus its cost.
Howard Rubin argues AI's fundamental limit may be economic rather than technological: it requires massive physical infrastructure like data centers, chips, and electricity, making intelligence a costly industrial product.
Organizations track AI usage well but struggle to measure real economic value, since "activity is not value."
Infrastructure costs are ultimately absorbed by consumers, enterprises, or governments even as per-unit computing gets cheaper, producing an "IT inflation" effect as total spending rises.
Rubin calls for a measurement revolution, including technology balance sheets and gross domestic intelligence metrics, to turn intelligence into measurable value.
Analytics Insight explains prompt engineering fundamentals, from prompt structure to evaluation in production.
This guide defines prompt engineering as crafting clear instructions -- covering task, context, constraints, output format, and quality criteria -- to guide AI models toward useful results.
It highlights the rise of 'context engineering,' which considers what information (instructions, documents, search results, tool outputs) should reach a model at each stage of a task.
The piece distinguishes zero-shot from few-shot prompting, noting results vary by model family, and observes that newer reasoning models often perform better with concise, direct prompts rather than explicit step-by-step instructions.
It also stresses that production prompts now need software-like rigor: version control and systematic evaluation across representative test cases.
AI technical debt is quietly accumulating across prompts, data, and models as enterprises scale pilots without governance.
AI technical debt builds up silently across prompts, data, models, and workflows as organizations scale AI pilots without proper governance, unlike traditional code and infrastructure debt.
It hides in distributed systems such as notebooks and departmental experiments, making it harder to track. Author David Linthicum argues enterprises are creating production dependencies without production discipline, turning successful pilots into unmaintainable systems with unclear ownership.
He recommends CIOs treat AI components as enterprise assets needing versioning and lifecycle management, establish clear production paths before scaling, and maintain an AI debt register. Winners will treat AI as long-term architecture, not a series of demos.
Laptop-class models have improved sharply, and this walks through model choice, parameters and runtime tuning in Ollama.
Local models have improved enough to warrant a second look: Posit engineer Simon P. Couch noted that models scoring 0% on his agentic coding eval a few months earlier hit 90% after the Qwen 3.5 and Gemma 4 releases.
They still will not match frontier cloud models, but for defined tasks like answering coding questions, writing functions or summarizing documents they can be surprisingly capable. Ollama is not the fastest option but is among the easiest to set up and is supported out of the box by VS Code, JetBrains AI Assistant, Zed and Posit Assistant, and can launch Claude Code or Codex against a local model.
Lasso Security finds SynthID-Text watermarking alters tool calling and safety refusals, especially under prompt injection.
The EU AI Act requires providers to mark model output with machine-readable code for provenance, and Google DeepMind's SynthID-Text, adopted by Anthropic and OpenAI, does this by nudging the model toward one statistically likely next token over another.
Lasso Security reports that because the technique changes the process by which each token is generated, it can also change safety behavior, including whether a model refuses a harmful request and whether that refusal holds under prompt injection.
The altered behavior is not necessarily worse, but it can be. Crucially the effect crosses organizational boundaries: an agent built on a third-party harness or an API client calling an Anthropic model inherits whatever output variation follows from Anthropic's watermarking.
ACM authors warn that AI-generated contributions are outpacing the volunteer maintainers who must vet them.
Six authors writing for the ACM Technology Policy Council, including Simson Garfinkel and Josiah Dykstra, argue that AI coding tools are making open source harder to maintain and secure.
The tools write code and find flaws quickly, but human maintainers still have to judge the output, and much of what now arrives is poor quality.
The funding asymmetry is stark: the Linux Foundation took in $292 million in 2024 while the Apache Software Foundation, cited as an example of volunteer-and-sponsorship projects, took in $2.4 million. AI-found flaws also push organizations to patch sooner, which is hard for applications assembled from many components that each need their own patch and re-release in sequence.
A Workera report finds 67.8% of employees now use AI regularly, but 56% get no work time allocated to upskill.
A new Workera 2026 State of Skills Intelligence Report finds 67.8% of employees now use AI tools beyond ChatGPT several days a week, up from 39.9% a year earlier, but 56.4% say no work time is allocated for upskilling and 42.5% cite a lack of learning materials.
Workera CEO Kian Katanforoosh urges leaders to carve out dedicated time for AI upskilling rather than assuming employees will figure it out alone, arguing that time also creates the psychological safety needed to experiment. The survey covered 1,000 salaried professionals at large US organizations.
AI ad agents can execute campaign changes directly, but clear rules must define where automation ends and human judgment begins.
Advertising platforms now let AI agents execute campaign changes directly rather than just recommend them, raising accountability questions when something goes wrong.
The article argues operational tasks like pausing over-budget campaigns suit autonomous execution, while strategic decisions still need human oversight.
Agents perform far better when given detailed context about goals and funnel structure, sometimes doubling conversion performance. Access typically runs through scoped, revocable API tokens rather than account credentials, but execution-level access requires more deliberate governance than read-only reporting access.
The author concludes success depends on setting clear limits and keeping human judgment close to decisions that still need it.
Comprehensive guide comparing 22 AI tools across seven categories for writing, coding, content creation, and productivity.
The article surveys 22 AI platforms organized by category including chatbots, writing tools, SEO platforms, note-taking apps, content creation suites, coding assistants, and voice helpers. Each tool is evaluated for specific use cases, with guidance that the best AI tool depends on the bottleneck you want to remove from your workflow.
The piece emphasizes selecting specialist tools for focused tasks versus general assistants for diverse workflows, noting that AI output requires human review for accuracy and sensitive work.
Many enterprise human-in-the-loop AI controls are performative, lacking real domain knowledge, authority, or override capability.
The article challenges organizations that claim human-in-the-loop safeguards on AI systems, arguing many such controls are performative. Genuine oversight requires reviewers with domain knowledge, actual decision authority, and the technical means to override or halt AI actions - not merely a screen to watch.
Decision fatigue and automation bias erode effectiveness over time, as reviewers begin rubber-stamping approvals once the AI seems reliable.
The piece also notes human oversight is not universally appropriate; time-sensitive cybersecurity responses may need autonomy. IT leaders are urged to demand evidence of context, independent judgment, and real override capability.
A foundational primer on how text becomes the discrete units models actually process, and why it matters for cost.
This primer explains how tokenization converts raw text into the discrete units a model processes mathematically, covering the main algorithms, their trade-offs and how to evaluate tokenizer behavior for a given corpus.
The practical relevance is broader than it first appears: tokenization determines cost, since billing is per token, and it explains a family of otherwise mysterious model behaviors around numbers, code and non-English text.
For teams whose non-English workloads cost more than expected, the tokenizer is usually the explanation.
Computerworld cites Gartner warning enterprises lack security readiness for internal and external AI risk.
Computerworld cites Gartner's Pete Shoard warning enterprises deploying AI tools aren't ready for the security risks involved, arguing organizations have kept doing 'the same thing' despite it not working.
Key risks include sensitive data leaking into LLMs without user awareness, hard-coded secrets pushed to public repositories via AI-assisted coding, and accidental exposure of files on public platforms.
The piece calls for a shift from reactive threat detection to proactive attack surface management that scans websites, repos, and social media for exposed assets, noting small and mid-sized businesses are especially vulnerable since they can't afford premium security vendors like Mandiant or CrowdStrike.
Gartner predicts AI inference costs per agentic workflow will rise more than fivefold by 2028 despite better token economics.
ARN reports Gartner's finding that improving token economics alone will not prevent overall AI costs from rising.
The analyst firm predicts AI inference costs per agentic workflow will increase more than fivefold by 2028.
The driver is organizations moving from assistive AI features and chatbots to increasingly complex, multistep agentic execution. Each step in an agentic workflow consumes tokens, so falling unit prices are outrun by rising consumption.
The finding is aimed at organizations budgeting AI on the assumption that model price declines will absorb growth.
Autonomous agents that generate code and probe infrastructure are exposing containers as a weak security boundary.
Containers won on packaging and delivery speed, but they were never designed as a hard security boundary, and autonomous AI agents are making that limitation obvious.
Agents dynamically generate code, install packages and probe surrounding infrastructure -- behavior that shared-kernel container isolation cannot reliably contain.
The industry response is a move toward microVMs such as Firecracker and Kata Containers, which keep container-style workflows while adding a separate kernel or hardware-enforced isolation.
The author argues no single boundary should be trusted alone, and recommends layering restricted networking, externalized credentials, governed API access and runtime monitoring.
AI-first mandates repeat the cloud-first mistake; organizations must be AI-enabled before they can responsibly be AI-first.
The author draws a direct parallel to cloud-first a decade ago: fast adoption, thin governance, and years spent unwinding security and cost problems, until OMB formally pivoted federal policy to cloud-smart in 2019, not to slow adoption but to make it sustainable.
The urgency around AI is real and sitting it out loses ground that will not come back, so AI-first as an instinct is correct.
The problem is sequence. Before going AI-first an organization has to be AI-enabled: sanctioned tools available, data structured for AI consumption, basic guardrails in place. Most ungoverned adoption today is organizations attempting AI-first while still only AI-enabled at best.
Autonomous AI agents create new security risks requiring identity controls, tool restrictions, and runtime monitoring.
Agentic AI systems that access tools, databases, and business platforms introduce expanded attack surfaces beyond traditional chatbot safety concerns.
Prompt injection attacks can escalate into system compromises when agents possess tool execution capabilities, as demonstrated by critical vulnerabilities in Microsoft's Semantic Kernel.
Organizations must implement layered defenses including identity management, least-privilege access, parameter validation, persistent memory integrity checks, runtime policy enforcement, and human oversight for high-risk actions to effectively contain agent-based threats.
When leading models approach a benchmark's ceiling, scores stop distinguishing meaningful capability differences.
Benchmark saturation occurs when leading systems approach a test's ceiling, at which point score differences stop carrying information about real capability gaps.
The article explains the mechanisms, including test set contamination and the way optimization pressure concentrates on whatever is measured.
It then covers practical evaluation controls: building private held-out sets, measuring on tasks that mirror actual workload, and treating public benchmark position as a filter rather than a decision.
For teams selecting models, it is a useful primer on why leaderboard rank correlates poorly with production results.
BigDATAwire argues analytics teams must shift from building dashboards to supplying the semantics AI agents consume.
BigDATAwire examines how AI is changing what analytics teams are for. For years those teams built dashboards, reports and pipelines, established definitions and documented metrics so business leaders could answer questions.
Employees across enterprises are now turning to ChatGPT, Claude, Copilot and other AI systems instead of opening dashboards directly.
The difference matters: a dashboard returns a metric, while an agent interprets information, combines data from multiple systems and applies logic. Author Soham Mazumdar argues that shifts the analytics team's job toward defining and governing the semantics agents rely on.
Adversaries can manipulate the reasoning of AI defensive tools to compromise networks without tripping alerts.
The article argues that AI governance has become an operational security requirement rather than a compliance exercise, because adversaries can manipulate the reasoning of AI-based defensive tools to move through a network silently.
When a detection system reasons about what it sees, that reasoning becomes an attack surface: a crafted input can persuade it that malicious activity is benign.
Traditional signature-based controls fail differently and more visibly. The practical implication is that AI security tooling needs adversarial testing of its decision logic, not just accuracy benchmarks on clean data.
More from Artificial Intelligence · All IT section newsletters · This week