The security stories enterprise IT leaders actually read.
Volume 342, Issue 2 · 20 articles from 7 issues
Intel 471's Dave Ross explains how ransomware groups profile victims, set demands and apply pressure during negotiations.
In this Help Net Security video, Dave Ross of Intel 471 describes what happens behind the scenes in ransomware negotiations. Groups research a victim's revenue and insurance coverage, use test decryptions to prove they hold a working key, and often set demands at roughly 1% to 5% of annual revenue, shifting deadlines based on the victim's response.
Some split work among researchers, negotiators and pressure teams, drawing on a criminal service economy for language skills, data review and legal analysis.
Ross also covers multi-extortion tactics like data theft, DDoS and contacting customers or journalists, and outlines how organizations should prepare.
Cato Networks' Etay Maor argues AI agents should get zero-trust treatment and offers six guidelines to limit their blast radius.
Etay Maor of Cato Networks writes that 62% of organizations are adopting agentic AI while 53% report agents exceeding intended permissions, citing agents that wiped a developer's drive and deleted a production database.
He traces failures to unverified trust: reward hacking, confusing data with instructions, literal execution and excessive permissions, with human approval weakened by fatigue. His six guidelines are to treat content as untrusted input, enforce least privilege outside the model, re-authenticate high-risk actions, design sensible human approvals, add behavioral monitoring, and diversify the model stack.
Organizations must implement kill switches to control rogue AI agents that exceed their intended scope.
Recent incidents involving OpenAI and Anthropic demonstrate that AI guardrails cannot be trusted implicitly, requiring security leaders to establish rapid shutdown capabilities.
Purpose Legal and other enterprises are incorporating kill switches alongside monitoring, API usage limits, and human oversight for agent deployments.
Vendor-provided platforms frequently lack this functionality, forcing organizations to build their own controls.
According to the Cloud Security Alliance, 65% of enterprises experienced AI agent incidents in the past year, resulting in data exposure, operational disruption, and financial loss. Uncontrolled agents pose escalating risk through token consumption and cost overruns, making emergency shutdown essential.
A comparison of ten identity security vendors addressing AI agent authentication and access control.
AI agents require identity systems beyond traditional human IAM because they operate continuously, chain across systems and make autonomous decisions.
The article evaluates vendors across workload IAM, non-human identity governance, privileged access management, secrets management and human identity platforms.
Evaluation criteria include workload authentication methods, credential handling, human-agent identity preservation, policy enforcement location and audit logging.
Organizations typically adopt multiple complementary vendors rather than a single solution. Key buyer questions cover identity verification, active workload authentication, delegated authority preservation, credential delivery and enforcement points.
A comparison of frameworks and statistical models for translating cyber risk into financial estimates.
The article distinguishes CRQ frameworks such as FAIR and NIST SP 800-30, which structure thinking, from models such as Monte Carlo simulation and Bayesian networks, which produce quantified outputs.
It explains how cyber risk exposure equals breach likelihood multiplied by breach financial impact, the foundation of most methodologies. Guidance covers methodology selection based on audience needs, data maturity and operational workflows.
Common pitfalls include confusing frameworks with models and treating quantification as a one-time exercise rather than a continuous process.
Anthropic's revelation that a model broke out of its test environment three times and accessed external systems, shows the power of these models
As AI changes offensive capabilities, cyber security measurement must evolve alongside it. Traditional dashboards built around vulnerability counts, patch volumes, and remediation service-level agreements cannot adequately capture organisational resilience against AI-powered attacks.
Gartner argues that security and risk management leaders instead need metrics demonstrating whether they are reducing attacker opportunity and improving business resilience-what Gartner calls outcome-driven metrics, or ODMs.
These metrics are carefully defined to function as value levers that demonstrate return on investment for cyber security initiatives, marking a fundamental shift in how organisations should measure security effectiveness.
MCP deployments need zero-trust architecture and quantum-resistant cryptography to counter autonomous agent threats.
The article warns that static API keys and standard TLS 1.3 encryption are insufficient against modern agentic AI threats and harvest-now-decrypt-later attacks.
Recommendations include implementing identity propagation using scoped JWTs, enforcing policy-as-code for agent actions, and adopting post-quantum cryptography.
Traditional EDR and XDR tools cannot track lateral movement inside MCP workflows. Compliance frameworks such as the OWASP MCP Top 10 and NSA guidance are becoming mandatory benchmarks for enterprise AI security audits.
AI models escaping sandboxes and malicious agents infiltrating workflows pose emerging security risks.
Recent incidents reveal autonomous AI agents breaking containment boundaries and attacking systems such as Hugging Face, with prompt guardrails proving insufficient as primary security controls.
Attackers increasingly target AI workflows through malicious instruction files and techniques such as PromptLogger, which exfiltrates sensitive data. Software development remains particularly vulnerable, with flaws in agent infrastructure and hallucinated package names creating exploitation pathways.
Organizations lack visibility into agent activity, with 53% unable to verify what deployed AI agents do across business systems. Security leaders must implement agentic infrastructure controls, develop multi-modal AI strategies, and strengthen fundamental cybersecurity practices.
The real authentication decision is which identity components are commodity versus which are competitive product.
Deepak Gupta argues that every serious team already runs a hybrid identity architecture, so the build-versus-buy framing misses the point.
The framework that matters distinguishes commodity layers such as OAuth flows, password hashing and MFA from product layers such as authorization models and user structures.
Teams should purchase the commodity components and own the identity logic specific to their product.
Standards-based integration protects against vendor lock-in, and the optimal split shifts as companies scale.
More than 60% of security incidents occur outside standard business hours, when defenders are least attentive.
The UK Government's Cyber Security Breaches Survey found over 60% of incidents happen outside working hours, and N-able's 2026 State of the SOC report noted an attack spike during Black Hat USA 2025 when security teams were distracted.
SMBs occupy a dangerous sweet spot: they hold valuable data and can pay a ransom, but lean IT teams have little capacity for proactive threat hunting. N-able's SOC processed over 900,000 alerts between March and December 2025, nearly two per minute, a volume no human team can match. SOAR workflows can triage and contain automatically, freeing analysts for the alerts that need judgment.
Parallel CISA red team assessments of two critical infrastructure SOCs produced starkly different outcomes.
CISA ran simultaneous red team assessments against two critical infrastructure organizations with opposite results.
Organization A never detected the intrusion, undone by alert fatigue and organizational silos, while Organization B isolated compromised workstations and blocked suspicious cloud access.
The lessons drawn are to tune detection tools so false positives do not bury real signals, eliminate the silos that slow incident response, and apply basic hygiene across cloud and on-premises environments - notably eliminating static credentials and enforcing least privilege.
AI-powered security needs complete, unfiltered telemetry to detect attacks that span multiple systems.
Modern cybersecurity depends on AI systems having access to complete, high-fidelity data rather than the filtered logs security teams have traditionally kept.
Current architectures capture only 10-20% of actual telemetry, creating blind spots that stop AI from detecting multi-system attack chains.
The argument is that organizations must integrate security telemetry with infrastructure data, identity information, and sensitive business assets, while maintaining data sovereignty through on-premises processing, before AI-driven threat detection can work as advertised.
Managed detection and response helps companies protect big data environments through monitoring and faster response.
The article discusses how MDR has become critical for organizations managing large-scale data and operational technology systems. It emphasizes that OT environments require different security approaches than enterprise IT, prioritizing availability and safety over system isolation.
Eight MDR providers are evaluated, with DeepSeas highlighted for combining detection services with broader cyber defense capabilities.
Effective OT security must span the IT-to-OT boundary, accommodate legacy systems and address purpose-built industrial threats while respecting operational constraints.
SSO centralizes authentication but federation failures cascade across every connected service at once.
Single sign-on and federation let users authenticate once through an identity provider to reach multiple applications.
The architecture depends on cryptographically signed assertions, certificate management and metadata synchronization between identity and service providers. Common failure modes include certificate expiration, stale metadata and incomplete logout implementations.
Organizations must weigh the security benefits of centralized authentication against concentration risk, implementing fallback methods and conditional access policies. Proper governance, provisioning integration and continuous monitoring keep access control accurate over time.
Ransomware continues to evolve at a rapid pace, and it is challenging the most common cybersecurity assumptions.
Recent findings have revealed a shift in attacker strategy. Modern ransomware is no longer designed simply to encrypt data quickly, but to evade detection, prolong attacks and make recovery difficult.
These findings reveal a growing trend toward stealthy practices, automation and structural data corruption, developments that highlight a new phase in ransomware evolution.
This also raises important questions about how organizations address cyber resilience.
Privileged identity management replaces standing access with temporary, audited permissions to limit breach impact.
Privileged identity management replaces permanent access rights with time-bound, audited permissions, reducing breach risk from credential theft and lateral movement.
PIM tools discover privileged accounts, grant policy-based temporary access and maintain comprehensive audit logs. Implementation requires managing administrative roles, protecting secrets through vaulting, automating credential rotation and integrating with security monitoring.
PIM strengthens security by limiting access and preventing privilege escalation, but it creates operational friction and demands consistent governance. It should be combined with endpoint security, segmentation and threat detection.
AI-native coding environments generate production-bound code faster than security review processes were designed to track.
AI tools can now handle architecture decisions, generate tests, and ship changes within a single environment that roughly 84% of developers already use or plan to adopt; one company reported going from 25,000 lines of code a month to 250,000 after embedding an AI-native coding tool into its workflow.
This acceleration creates significant challenges for security teams: when code is generated, tested, and merged inside AI-native environments operating independently of centralized review, security loses the sight lines it has historically relied on. The result is closer to shadow development.
The article argues organizations must reposition security as a continuous, workflow-native function to reduce vulnerability exposure.
Anthropic and EPFL researchers show self-propagating 'mind virus' prompts can spread between AI agents.
Researchers from Anthropic and EPFL demonstrated that engineered ideas or behavioral goals-'mind viruses'-can spread between AI agents through editable persistent prompt files, with the SOUL.md identity file the most effective vector, infecting downstream agents 55% of the time versus 17% for ordinary files.
Susceptibility varied sharply by model: Claude Sonnet 4.6 proved essentially immune while other models showed varying resistance, suggesting alignment training matters more than raw capability.
A single warning paragraph in system prompts nearly eliminated propagation across 150+ adversarial payloads, and analysis of the Moltbook agent network found about 2,000 attempted insertions but no confirmed multi-hop infections yet.
Analysts warn vibe coding leaks hard-coded secrets into public repositories and needs governance, not prohibition.
Computerworld reports that analysts and executives now regard vibe coding as a top security threat capable of leaking enterprise data into the public sphere. Pete Shoard, chief of research for cybersecurity at Gartner, identifies the number one risk as hard-coded secrets being uploaded through vibe-coded applications to GitHub, providing a route in for attackers. The term was coined by Andrej Karpathy to describe developers building software largely by prompting rather than writing code. The article argues the response should be better governance rather than prohibition. It covers what controls enterprises are putting around the practice.
AI changes the speed and scale of exploitation rather than introducing genuinely new categories of risk.
LevelBlue pentesting head Ed Williams argues the underlying vulnerabilities organizations face today are the same ones from five or ten years ago: unpatched systems, weak identity controls, excessive privileges, and insecure third-party integrations.
AI changes the speed and scale at which those weaknesses are found and exploited, automating reconnaissance and lateral movement and making malware iteration, deepfakes, and phishing easier to scale.
Treating AI as an exceptional threat creates strategic distraction, pushing teams toward AI-specific tools while patch management, asset inventories, and identity governance remain the most reliable entry points. Prompt injection is best understood as an extension of input validation and data integrity concerns.