The security stories enterprise IT leaders actually read.
Volume 341, Issue 4 · 20 articles from 4 issues
CISOs should establish incident notification protocols before crises, prioritizing harm reduction over polished narratives.
Security leaders must prepare customer notification strategies during peacetime rather than amid active incidents, according to Zscaler CISO Sam Curry.
Key principles include letting harm reduction drive timing decisions, treating legal obligations as forcing functions, and giving customers actionable intelligence rather than a polished narrative.
Decision rights, escalation paths, and communication cadences should be established in advance through cross-functional work with legal, communications, and business leaders. The approach emphasizes tight facts with clear caveats and empathetic professionalism over overconfident statements.
Organizations deploy AI systems while lacking full visibility into how they operate, creating a trust-versus-opacity security paradox.
The article addresses a paradox in enterprise AI adoption: organizations are trusting and deploying AI systems despite lacking complete visibility into how those systems actually work.
It frames this visibility gap as a core security concern, questioning why businesses place trust in opaque, black-box AI mechanisms.
The piece highlights the security implications of operating AI without full transparency. (The full body was not retrievable from the fetched page, so supporting detail and expert commentary could not be captured.)
AI agents create distinct vulnerabilities through autonomous action, broad permissions and deep system integration.
AI agents introduce four primary risk categories: input manipulation attacks such as prompt injection, execution risks from over-permissioned access, architectural vulnerabilities in multi-agent systems, and governance blind spots.
Unlike standard language models, agents autonomously execute actions across databases and APIs, expanding the attack surface significantly.
Organizations should implement layered guardrails, quantified risk visibility and continuous monitoring before scaling agentic deployments into production workflows.
Unauthorized AI tools and embedded AI features operating outside oversight create significant enterprise risk.
Employees adopt AI applications faster than organizations can establish governance frameworks, producing shadow AI risk.
The primary concern is unauthorized data movement rather than unauthorized software alone, with 47% of enterprise generative AI usage occurring through personal accounts.
Security teams face visibility gaps as AI capabilities arrive silently through software updates and employee-built workflows. Organizations need continuous monitoring and employee training to understand where AI operates and what sensitive data it can reach, treating governance as ongoing rather than annual.
A new report puts global cybercrime damage in the trillions annually, with roughly $10,000 lost per incident.
A recent report highlights the escalating financial impact of cybercrime worldwide, with victims suffering substantial losses per incident. The data indicates individual victims lose an average of nearly $10,000 in each attack.
The cumulative effect represents trillions of dollars in damages annually across the global economy. The figures underscore the growing threat posed by cybercriminals and the need for stronger security measures and awareness.
Seven AI security risks in 2026 span data poisoning, prompt injection, model theft, shadow AI, and more.
The article catalogs seven top AI security risks for 2026. These include data poisoning that corrupts training sets, prompt injection (ranked OWASP's top LLM vulnerability), and model theft via systematic querying of public interfaces.
It also covers shadow AI, noting 15% of employees paste company data into chatbots, and supply chain vulnerabilities exemplified by the Ultralytics YOLO cryptominer incident.
Rounding out the list are sensitive data exposure from model memorization (as with Samsung and ChatGPT) and adversarial attacks that cause misclassification in systems like autonomous vehicles and security detection tools.
Outlines five key security risks of giving autonomous AI agents access to enterprise systems and how to mitigate them.
The article identifies five critical security risks when deploying autonomous AI agents with access to enterprise systems.
Primary concerns include granting agents excessive permissions without tool-level enforcement, inadequate user verification that enables impersonation, and reliance on long-lived credentials that increase breach impact.
Organizations often lack runtime enforcement controls, depending instead on prompts as security boundaries, an approach proven unreliable. The piece warns that prompt injection and credential compromise can turn routine agent operations into unauthorized business actions, recommending short-lived tokens, just-in-time access, and centralized policy enforcement as mitigations.
Organizations lack visibility into AI usage across employee tools and internal applications, creating security gaps.
FireTail addresses the AI visibility problem through two collection pipelines: workforce observability for employee-used tools and workload observability for internal AI systems.
Existing security tools cannot detect AI activity because it operates outside traditional file and network monitoring channels.
The platform normalizes provider data into a single audit trail so security, compliance and finance teams can track usage, detect sensitive data exposure and control costs.
A unified record provides defensible answers for audits and regulatory compliance while revealing shadow AI usage.
As AI adoption accelerates, attackers exploit the same technology, forcing teams to defend both traditional infrastructure and AI systems themselves.
The article identifies six critical AI security threats: prompt injection attacks that manipulate LLM behavior, AI-generated phishing and deepfakes that raise social engineering success rates, data poisoning that corrupts model training, AI agent vulnerabilities exposed through API and database access, and shadow AI where employees use unapproved tools and risk data exposure.
Unlike conventional attacks targeting infrastructure, these threats exploit prompts, training data, and automated workflows. Organizations need AI governance policies, red-team testing, production monitoring, and employee education.
The shift requires treating AI systems as a critical security perimeter alongside networks and endpoints.
Defense in depth requires multiple coordinated layers across the data lifecycle to protect sensitive information.
A single security control leaves exploitable gaps, according to software engineer Venkata Pavan Kumar Gummadi.
The strategy involves four coordinated layers: classification labels for sensitive fields, tokenization to replace high-value data, data loss prevention policies that monitor data paths, and access controls using identity-based short-lived credentials.
A case study walks through how each layer catches threats the previous one missed, ultimately enabling recovery from a ransomware attack.
Cybersecurity approaches need fundamental transformation to address emerging threats from AI agents.
The article examines how AI agents are reshaping the cybersecurity landscape and forcing updated defense strategies.
Traditional security measures prove insufficient against AI-driven threats that operate with increasing sophistication and autonomy.
Organizations must adapt their approaches to protect against these evolving challenges. The piece emphasizes the urgency of developing new frameworks and technologies capable of defending against AI-enabled attacks while maintaining operational efficiency.
Modern ransomware attacks target mid-level managers rather than executives to accelerate payment decisions.
Zscaler researchers analyzed 351 victims across 334 organizations in a single ransomware campaign and found nearly two-thirds held manager-level positions or above.
The average victim was a 46-year-old Gen X employee, with three-quarters working in accounting, finance, sales, operations, HR, or marketing. Attackers now combine compromised system data with public information to map organizational structures and identify individuals with business authority.
Rather than pursuing technical privilege through administrator accounts, criminals target business privilege, meaning employees who control payments, budgets, contracts, and vendor relationships. The shift reflects a sophisticated, targeted extortion approach replacing indiscriminate attacks.
Average breach costs reached $6 million globally, up 35% year over year, with AI-enabled attacks driving the rise.
IBM's 2026 Cost of a Data Breach report found that average breach expenses rose to $6 million from $4.44 million previously, based on analysis of 600 organizations worldwide.
One in four malicious breaches involved AI-enabled attacks, particularly deepfakes and AI-driven malware.
While security automation and AI tools reduced breach costs by nearly $2 million on average, one quarter of organizations have yet to adopt these defenses.
The US experienced record costs at $11.5 million per breach, driven by regulatory penalties and business expenses, nearly double the global average.
Traditional annual pentests are obsolete as AI-speed threats drive a shift to continuous, programmatic testing.
The article argues that the traditional point-in-time annual penetration test is effectively dead, unable to keep pace with a machine-speed threat landscape where exploits emerge in minutes rather than months.
While periodic pentests persist for compliance, insurance, and contractual reasons, these represent a compliance floor rather than genuine security. It notes practitioner disillusionment with autonomous pentesting tools that miss critical vulnerabilities and produce false negatives.
The piece advocates evolving toward continuous, AI-enhanced testing and exposure management that feeds actionable results into broader vulnerability management programs. (Page returned HTTP 403; details reconstructed via search.)
Zero trust treats identity as the primary boundary, requiring continuous verification before granting resource access.
Zero trust replaces perimeter-based security by verifying every access request using identity attributes, device posture and contextual signals.
Unlike traditional models that grant broad access after initial authentication, it minimizes breach impact through least-privilege permissions and continuous monitoring.
Implementation requires five capabilities: a policy engine, device trust, identity verification, an access proxy and comprehensive logging.
Organizations should deploy incrementally, prioritizing by business criticality while managing the operational burden of policy creation across applications and user populations.
Static calendar-based security training fails; responsive, threat-driven human risk management builds real resilience.
The article argues that legacy security awareness training built on mandatory monthly videos breeds resentment rather than resilience and cannot match AI-driven attacks that ignore quarterly training cycles.
It advocates responsive Human Risk Management with three mechanics: threat-to-sim workflows converting real external threats into immediate simulations, micro-coaching delivered at the moment of failure, and role-specific multi-channel training.
The goal is to close the intelligence loop by turning current threat intel into contextual simulations before attackers exploit gaps. This shifts security culture from compliance burden to empowering business function.
AI SBOMs are needed as a practical control, not a compliance nicety, after the Hugging Face-OpenAI incident.
AvePoint's Dana Simberkoff argues the Hugging Face-OpenAI incident exposed the need for more ubiquitous AI software bills of materials.
SBOMs became standard after Log4Shell and a 2021 executive order, but AI changes the inventory problem: an AI system is a full stack of models, datasets, prompts, retrieval sources, APIs, plugins, identities, permissions, and agentic workflows that can act with limited human review.
The 2026 State of AI Report found 88.4% of organizations had at least one agent-related security incident in the past year, and 21.1% do not know whether employees use unsanctioned tools to build agents.
AI SBOMs must be living, machine-readable inventories refreshed through CI/CD and MLOps pipelines, not assembled after an incident.
The UK's NCSC releases guidance for organizations responding to highly disruptive cyber attacks.
The National Cyber Security Centre published a 31-page guidance document addressing organizational response to highly disruptive cyber attacks across three phases: immediate containment in the first hours, recovery over days and weeks, and organizational rebuild.
The guidance emphasizes that leaders should set a calm, professional tone from the outset to avoid rushing to judgement while supporting staff through uncertainty.
Graeme Gordon, CEO of Converged Solutions Group, notes that organizations should actively rehearse and test incident response plans rather than simply keeping guidance on file, recommending staff training, incident response team formation, and fire drills across attack scenarios.
Mature identity programs still fail because ungoverned machine identities operate undetected.
Machine identities now outnumber human users by as much as 50 to 1, yet most lack lifecycle controls and clear ownership. The 2026 Data and Identity Security Report found 43% breach rates among organizations where AI expanded identity counts.
The Drift OAuth token breach illustrated how trusted but forgotten credentials enable lateral movement across systems.
Organizations should establish current inventories, defined ownership, rotation schedules and automated deprovisioning for all machine identities before deploying AI agents at scale.
Retired NSA chief Paul Nakasone urges disconnecting PLCs after suspected Iranian attacks on US water systems.
At least 12 US states' water systems have been compromised in suspected Iranian cyberattacks targeting programmable logic controllers that monitor sensor data and control pumps.
Retired General Paul Nakasone, former NSA chief, said these PLCs should not be connected to the internet and called for higher standards in critical infrastructure defense.
US water infrastructure presents a massive attack surface across 50,000 municipalities, most underfunded with limited IT staff and no dedicated cybersecurity personnel.
Researchers suspect Iranian involvement based on historical targeting patterns, though no official attribution has been made. Nakasone advocates collaborative defense approaches including initiatives like DEF CON Franklin, where volunteers help secure water facilities.