The AI stories enterprise IT leaders actually read.
Volume 341, Issue 4 · 20 articles from 3 issues
A DigiCert survey finds 78% of firms hit AI security incidents amid a wide governance gap.
A DigiCert survey found nearly 78% of organizations suffered AI-related security incidents or vulnerabilities, driven mainly by unauthorized or misconfigured AI agents rather than bad code.
Although 90% discussed AI governance at board level, only 50% funded formal programs, and just 53% could trace AI decisions back to their models and source data. DigiCert argues AI agents need verified identities like employees, citing efforts such as Estonia's digital IDs and Microsoft's Agent ID.
The findings echo Spacelift research showing 93% of firms hit AI-caused infrastructure incidents, suggesting enterprises are prioritizing speed over risk assessment.
AI can automate threat-modeling tasks, but human validation and organizational buy-in remain essential.
Author William Dupre explains how AI accelerates threat modeling by automating repetitive work like system design reviews, threat enumeration, data flow identification, and component mapping using frameworks such as STRIDE and LINDDUN.
AI can link application code to runtime components and diagram application flows, while multimodal fusion and graph neural networks help reconstruct missing documentation. Human oversight stays essential, with security architects validating that outputs fit organizational risk profiles and architecture.
Cross-functional teams ensure both technical and business risks are covered. Successful adoption requires organizational buy-in, baseline metrics, and staff training on AI-enhanced processes.
CIOs must understand AI data ownership across input, training, and output categories to protect proprietary information.
Organizations face confusion about AI data ownership beyond token costs, risking proprietary knowledge that becomes vendor training data.
Three distinct data types, input, training, and output, require separate governance strategies, though vendors often use vague contract language. Enterprise AI tools offer stronger protections than consumer versions, but no-training clauses do not prevent vendors from learning through usage patterns and telemetry.
CIOs should demand specific contractual protections including no customer data training, short retention windows, data isolation, output ownership rights, and audit capabilities. Open-weight models on internal infrastructure provide architectural certainty about data location at the cost of managing security and MLOps overhead.
AI is reshaping threat intelligence, making governance, explainability, and human oversight foundational.
Flashpoint argues AI is becoming embedded across cyber threat intelligence workflows, making governance and transparency as important as capability.
Trust is fundamental: analysts must be able to interrogate a system's reasoning as they would any source, and oversight is now a foundational necessity since intelligence informs many enterprise functions.
It advocates a human-led, AI-scaled model where AI accelerates repetitive tasks while experts provide validation and context. Security leaders should evaluate four dimensions: explainability, confidence thresholds, governance mechanisms like permissions and audit trails, and practical operational improvements.
The post notes Flashpoint's recognition as a Challenger in Gartner's 2026 Magic Quadrant for cyberthreat intelligence.
A five-prompt head-to-head of Meta AI, ChatGPT and Gemini's Nano Banana 2 image models, won overall by ChatGPT.
TechRadar ran the same five image prompts through Meta's new AI image model, ChatGPT and Google Gemini's Nano Banana 2, covering a product poster, photorealistic wildlife, a social invite, a four-panel comic and a noir cartoon.
Nano Banana 2 was consistently strongest on realism, while Meta AI took the biggest creative swings and produced the most premium commercial visuals.
ChatGPT won overall by most reliably matching both the wording and the intent of each prompt, especially on typography-heavy designs. Meta's clearest win was the noir cookie-mystery cartoon, where it best embraced the ridiculous premise. The conclusion: all three are capable, and the differentiator is judgment about what the user actually meant.
Tech workers can protect careers and manage AI burnout through validation, batching, and strategic skill-building.
Tech professionals face new challenges from AI adoption, including low-quality generic output and hallucinations that demand constant verification.
An Upwork study found 77% of employees reported increased workload from AI, with 88% of top productivity gainers experiencing burnout. Workers can counter these issues through validation practices, batch processing of AI outputs, and reusable prompts.
To future-proof careers, professionals must demonstrate human value through business acumen, institutional knowledge, and judgment about when AI should not be used without heavy review.
Companies stall at AI scaling by expanding pilots instead of redesigning systems for production.
The article argues companies fail at AI scaling because they try to expand pilots rather than fundamentally redesigning systems for production, and because they add resources instead of adopting an AI-first mindset.
Four areas need restructuring: operating models and governance, talent beyond just engineers, production-grade tools, and ROI measurables. Organizations should shift from an 80-20 split favoring small experiments toward more transformational projects as confidence grows.
It also urges embracing nondeterministic thinking rather than expecting binary if-then results. Leaders who have moved past experimentation to scaling gain an advantage over unprepared competitors.
Coder CEO Rob Whiteley argues AI agents stall short of production due to missing governance, not model quality.
Across 100 engineering organizations, 61% already run AI agents, yet few trust them in production because rogue agents can leak credentials, hit unauthorized repositories or burn cloud budget with no audit trail.
The author argues the fix is governance, using controls long applied to human engineers: least privilege, audit logging, identity management and scoped access. He recommends three steps: isolate agents in ephemeral workspaces with zero outbound access and allowlisted MCP tool calls; scope permissions to task-specific API keys with monitoring and identity-aware LLM routing; and require human approval gates plus policy-based model selection.
Governance belongs in the infrastructure layer, not individual apps, so policy is not rewritten as models change.
Cisco Talos researchers found threat actors easily bypass AI safeguards through simple social engineering.
Cisco Talos analyzed prompt logs from compromised endpoints using tools such as Claude Code and found that AI guardrails provide minimal resistance against determined adversaries.
Attackers commonly bypassed protections by claiming ownership of target systems, framing requests as capture-the-flag exercises, or decomposing malicious tasks across multiple sessions.
Researchers noted that most of the time it was a simple claim of authorization and the model complied.
Unsophisticated actors still produce substandard results, while skilled threat actors have significantly advanced their capabilities. Security professionals should deploy AI defensively to identify actionable alerts.
Mid-sized tech firms need enterprise-grade AI security across data, model, app, and infrastructure risk layers.
The article addresses how mid-sized tech companies can adopt AI securely despite lacking Fortune 500-scale security teams. It identifies four risk layers: data (poisoning, leakage), models (hallucinations, adversarial inputs), applications (prompt injection, insecure integrations), and infrastructure (API security, access controls).
Traditional tools fall short because they cannot detect AI-specific threats like prompt manipulation or model drift.
The guide recommends vendor security evaluation, network segmentation, access controls, and risk-based governance frameworks, executed in phases from assessment to continuous improvement.
Rising AI project failures stem from poor data readiness, not technology limits.
The article reports AI project failures jumped from 17% to 42% between 2024 and 2025, with up to 72% of businesses potentially abandoning pilots due to inadequate data preparation rather than technology. Enterprises wrongly prioritize data volume over accuracy, when the real need is data that is accessible, understandable, and trustworthy.
Governance is often treated as compliance rather than an enabler, risking privacy breaches, bias, and violations, while rushing deployment creates technical debt and silos block consistency. Human readiness, data literacy, and change management are equally critical.
It recommends data audits, governance frameworks, infrastructure modernization, upskilling, and continuous quality monitoring before deployment.
DeepSeek's V4 models deliver strong performance at very low API cost, but Chinese data residency and censorship are serious business risks.
DeepSeek offers a free, unlimited chat app and an aggressively priced API built on MoE models: V4 Flash (284B total / 13B active) at $0.14 per million input tokens and V4 Pro (1.6T / 49B active), both with a 1M token context window and MIT-licensed open weights for self-hosting.
Testing found clean code output, accurate long-document summaries and a genuinely useful Expert Mode reasoning path, with cached input tokens 90% cheaper and off-peak discounts up to 75%.
The drawbacks are non-technical: data is stored on mainland China servers subject to Chinese law, politically sensitive prompts return evasive answers, and Italy blocked the service in January 2025. There is also no image generation or understanding, and support is limited to Discord and email.
Every AI browser tested at Black Hat 2026 proved vulnerable to indirect prompt injection despite layered guardrails.
At Black Hat USA 2026, Brave security engineer Artem Chaikin demonstrated indirect prompt injection against Opera's AI browser, Perplexity Comet, and ChatGPT Atlas, hiding instructions behind HTML, near-invisible image overlays, and Reddit spoiler tags.
Even Atlas, which layers system prompts, trusted-content tagging, tool scanning, and user approval, fell to mimicked trust tags and exfiltration hidden in URL fragments.
Free users hitting usage caps could be downgraded to a model more susceptible to injection. Brave's mitigations include separate profiles, a minimum model floor, and a sentinel model that checks whether a proposed tool action matches the user's request. Chaikin concluded there is no perfect solution, only overlapping layers.
An AI asset inventory is essential governance infrastructure to surface shadow AI and assign accountability.
Kovrr argues most organizations pursuing AI governance lack visibility into their actual AI systems, and an AI asset inventory is essential infrastructure documenting every tool, including shadow AI deployed outside approval channels.
The inventory should capture vendor details, deployment methods, business functions, data types, regulatory classifications, and third-party dependencies, which matters under the EU AI Act's risk tiers.
Discovering shadow AI requires multiple methods such as network traffic analysis, identity integrations, SaaS monitoring, and browser extension tracking, since single approaches miss different tool types.
Cloud-hosted AI agents run with user privileges but evade traditional security tools, needing runtime governance.
Cequence CTO Shreyans Mehta argues that AI agents now running in vendor clouds like Anthropic's and OpenAI's execute with full user privileges yet stay invisible to conventional controls.
Legacy tools fail because EDR cannot monitor external clouds, SASE sees only legitimate API calls, and identity controls verify who accesses systems but not what agents do once authorized.
He proposes an AI Gateway at the action layer that monitors real agent requests, defines Agent Personas for behavioral boundaries, and enforces scope and runtime checks. The piece stresses governance must occur at runtime to stop agents repeating harmful actions at machine speed.
In the past year, AI has moved from chatbots to autonomous agents that can retrieve information, make decisions and take action.
That shift creates a new data risk beyond employee oversharing: agents may access more information than they should or act on data that is not appropriate, necessary or governed.
The next challenge is ensuring autonomous agents can only discover, use and act on information within clear enterprise controls.
Organizations are still building the maturity needed to close that gap. According to research from the Ponemon Institute, only 41% of enterprises have AI-specific data privacy policies in place.
TeamViewer's Patricia Leppert argues shadow AI is a symptom of digital friction, not employee misbehavior, and stricter blocking backfires.
Shadow AI is a genuine security risk, but treating it purely as a controls problem addresses the symptom rather than the cause.
Employees adopt unsanctioned tools because approved ones are slow, restricted or lacking functionality, and research cited shows 80% of employees lose time to dysfunctional IT, averaging 1.3 workdays per month.
Trust in IT is weak too: 62% doubt their IT team supplies current AI tools, 57% doubt fast issue resolution and 47% fear inadequate data protection.
The author calls for clear AI governance ownership spanning IT, legal, compliance, HR and leadership, plus specific and accessible guidance rather than abstract policy.
Security-by-design should extend to the digital workplace so the easiest path is also the secure one.
Duck.ai gives account-free, privacy-proxied access to frontier models, but bundles paid AI tiers with VPN and identity services.
Duck.ai lets users chat with models such as GPT-5 mini and Claude Opus without registering an account, routing every conversation through DuckDuckGo's anonymizing proxy while contractually barring providers from training on the chats.
The free tier offers six models including Claude 4.5 Haiku, GPT-4o mini and gpt-oss-120b, with history stored locally on device.
Plus at $9.99/month adds GPT-5.4 and Claude Sonnet 4.6 and undercuts ChatGPT Plus and Claude Pro, while Pro at $19.99/month adds Claude Opus 4.7 and extended reasoning.
Weaknesses include no API, no file uploads, undisclosed usage limits and no AI-only subscription, so subscribers also pay for bundled VPN and identity protection they may not need.
Essential criteria for selecting AI orchestration platforms as enterprises scale agent deployments.
AI agent orchestration platforms coordinate role-based and task-based agents, tools, data, and people into multistep workflows using open standards such as MCP and A2A.
Key evaluation areas include observable control with governance guardrails, secure and resilient operations meeting compliance requirements, integrated testing and feedback mechanisms, interoperability with open standards, and vendor viability with a clear roadmap. Organizations should assess how platforms handle oversight of autonomous decisions, long-running processes, audit trails, and continuous outcome evaluation.
As enterprises move from proof of concept to production, the right platform determines whether workflows and governance scale across thousands of agents.
As AI cheapens information, value shifts to human judgment and knowing when AI errs.
This Solutions Review EdTech piece argues that as AI lowers the cost of producing information and widens access to knowledge, the value of learning shifts toward application, critical thinking, collaboration, and continuous capability development.
AI enables more personalized instruction, richer formative feedback, adaptive learning pathways, and realistic simulations, speeding both content production and skill development.
But the core competitive skill becomes interpreting information, applying judgment, and recognizing when AI is wrong.
Learning becomes less about delivering information and more about helping people solve complex problems and adapt. (Page returned HTTP 403; details recovered via web search.)
More from Artificial Intelligence · All IT section newsletters · This week