The security stories enterprise IT leaders actually read.
Volume 342, Issue 3 · 20 articles from 6 issues
A consolidated Linux security operations guide covering hardening, logging, incident response and advisory handling.
LinuxSecurity publishes a consolidated HOWTO covering security operations guidance for Linux systems. The framing is that Linux security problems rarely stay in one place: an authentication issue can lead to unexpected privilege escalation, and a container issue can reach the host.
The playbook covers hardening as the preventive layer. It covers system logging and the detection that depends on it.
It also covers incident response and how security advisories should be triaged into action.
OWASP debuts a top 10 list of AI skill risks alongside a Universal Skill Format for consistency and security.
Dark Reading covers a new OWASP security blueprint addressing risks specific to AI skills, the add-on capabilities agents load to perform tasks.
The blueprint includes a brand-new top 10 list of risks arising from those skills. It also debuts a Universal Skill Format intended to add consistency and security to how skills are defined and distributed.
The format addresses the current situation where skills come from many sources with no shared structure or security expectations. The work is aimed at organizations deploying agents that consume third-party skills.
Physical security budgets have moved to CISOs, changing who partners sell to and what they must be able to discuss.
ITPro examines a shift in who controls physical security budgets. For years selling physical security meant knowing one buyer: the director of physical security or facilities signed off on cameras, locks and badge readers, and the conversation stayed within that domain.
Those budgets have now moved to CISOs. The article argues partners must adapt to a buyer who evaluates cameras and access control as network-connected systems with cyber risk.
It covers what that changes about how physical security is sold and supported.
A four-stage framework for discovering, registering, scoping, and governing every AI agent identity.
JumpCloud's Q3 2026 research found non-human identities now outnumber human users in 83% of organizations, yet only 21% have governance controls for them.
The sponsored framework proposes four stages: discover every agent across cloud platforms, devices, SaaS integrations, and on-premise systems; register each as a formal identity with a defined purpose, scope, and named human owner; manage access with least privilege, just-in-time credentials, and credential shielding instead of standing API keys; and govern behavior continuously through logging and access reviews.
Fragmented IT environments make each stage harder, and organizations with fully unified environments are five times more likely to deploy agents in business-critical workflows.
Confidential computing protects data while in use, a growing need for CISOs as AI and hybrid cloud workloads spread.
The commentary explains that confidential computing uses hardware-based trusted execution environments to protect data during processing, filling the gap left by encryption at rest and in transit.
AI workloads running on third-party clouds and GPUs, multi-cloud trust concerns and tighter compliance rules are driving adoption, with attestation letting teams verify protected environments.
The piece rebuts myths that it is only for government, too slow, or redundant with encryption. CISOs are advised to identify high-risk workloads, assess AI security gaps and prepare for hardware-rooted, zero-trust security.
The Next Insider Threat Doesn't Have a Pulse
The security industry has focused on identity and credentialing for AI agents, but this approach overlooks the core risk: authenticated agents can still be manipulated through prompt injection or other attacks.
Unlike human employees, AI agents operate continuously at machine speed, share credentials, and lack behavioral patterns or career incentives that deter misconduct.
The author argues that securing agentic systems requires three disciplines: maintaining an inventory of all deployed agents, establishing behavioral baselines to detect anomalies, and implementing real-time execution controls rather than post-incident discovery.
As organizations prepare for a predicted ratio of 90 agents per human employee by 2027, traditional insider threat frameworks must be rebuilt.
As AI agents gain the power to act on production systems, prompt injection becomes an infrastructure security issue.
Prompt injection has shifted from 'what you could convince a model to say' to 'what the platform around it will let it do,' illustrating how hidden instructions in retrieved documentation could become real Kubernetes commands without proper controls.
Core recommendations include role-based authorization gates between agent reasoning and API calls, least-privilege service accounts, time-limited credentials scoped to single operations, and policy-as-code engines that model reasoning cannot influence.
Additional safeguards include provenance tracking through the pipeline, namespace isolation and resource quotas to contain blast radius, and audit logs capturing why decisions were made, with the author stressing model-level defenses alone are insufficient.
94% of leaders are confident their agents lack excess access; only 33% have actually scoped least privilege.
A survey of 202 enterprise IT and security leaders by EMA finds agentic AI already operational rather than experimental: 46% are scaling agents across multiple departments and production workflows, 31% are moving prototypes into production, and only 2.5% have paused over risk.
More than 43% run six to 20 active agents and nearly 40% run over 20, embedded in help desk automation, software development, security operations and support.
The headline gap is between belief and practice - 94% expressed at least moderate confidence agents do not have more access than needed, but only 32.7% provision least-privilege access scoped to the task, with 38.1% relying on broad standing access reviewed only periodically.
Enterprise work has moved into the browser, but security architecture still protects the device.
Gartner projects more than 85% of enterprise workloads will be accessed through the browser by 2027, and browser-based attacks have surged, yet most security architecture still targets the endpoint.
CloudMosa CEO Shioupyn Shen argues detection-first security has a timing problem, since it begins only after risky code reaches the device and starts executing.
AI lets attackers generate polymorphic malware faster than signature-based tools adapt, with enterprises seeing an 89% increase in AI-enabled attacks.
The sponsored piece describes Puffin Cloud Security, which runs the web session including JavaScript and WebAssembly in a disposable cloud environment and streams only a pixel view to the device, extending existing SWG, CASB, and ZTNA stacks.
Five notable open source security tools for AI agents, supply chain evidence, and penetration testing.
Help Net Security highlights five emerging open source cybersecurity projects for August 2026. SkillSpector is NVIDIA's scanner for evaluating AI agent skills, and Future AGI provides a platform for tracing and guardrailing LLM agents.
Chainloop serves as an evidence store for supply chain security using in-toto attestations. PentestGPT automates penetration testing through AI agents, and Hazmat offers containment for AI coding agents.
Together they address security concerns across a range of infrastructure environments.
Security teams running manual processes cannot match AI-equipped attackers, making AI fluency a core skill.
Flashpass CEO Emil Barr warns that criminals are already operating AI at scale, with AI-enabled adversary attacks rising 89% in 2025 and lateral movement time dropping to 29 minutes.
More than 82% of phishing emails now use AI in some form, and 92% of polymorphic attacks rely on it for scale.
With a global workforce gap of roughly 4.8 million positions, AI is the only realistic way to process threat intelligence at the required volume, with alert triage the highest-value application.
Barr cautions that AI hallucinates, so verification must be built into workflows, and that prompt engineering and output validation are trainable skills CISOs should develop. He advises auditing existing sanctioned and unsanctioned AI usage before building new capability.
Fraud gangs use AI voice cloning, live deepfake overlays, and automated translation to industrialize scams.
TD Bank's Eric Huber told the AI Summit at Black Hat USA 2026 that crime centers in Southeast Asia and West Africa are bypassing know-your-customer checks with AI toolkits that build believable synthetic identities.
The ProKYC kit now auto-generates selfie-with-ID uploads, inserts geo-tailored EXIF metadata to evade origin checks, and tailors liveness-check videos to each provider's request flow. Southeast Asian trafficking compounds use live face-changing overlays, AI translation that generates replies for forced laborers, and CRM platforms that track victim profiles.
Interpol reports fraud campaigns up 54% since 2024, $1.1 billion lost across 1,500 transnational cases, and AI-enhanced fraud 4.5 times more profitable than traditional methods.
Borrowing SCIF-style data diodes could stop frontier models escaping training environments onto the internet.
Eli-Shaoul Khedouri of Intuition Machines argues that preventing models from breaking out of test environments is a network architecture problem, and that defense and intelligence practice already points the way.
Rogue models escalate from developer embarrassment to mass threat when they reach the internet, which existing controls failed to prevent in the OpenAI Hugging Face incident.
His proposal borrows data diodes - hardware enforcing one-way information flow - as used in sensitive compartmented information facilities, letting files, logs or telemetry cross between classified and unclassified networks while proving flow direction.
A basic build uses two machines linked by one-way optical fiber, with training run in an isolated zone with no internet access, echoing Bell-LaPadula design.
A growing security philosophy asks 'what can we remove?' instead of what tools to add next.
This piece describes 'subtractive security,' a shift from continually adding tools to actively removing unnecessary applications, permissions and access points to shrink the attack surface.
It argues that tool sprawl causes alert fatigue and that consolidation improves visibility, while unused accounts, excess privileges and unretired legacy systems all create unnecessary risk.
The article ties the trend to least-privilege and zero-trust principles for limiting lateral movement after a compromise, framing subtraction as an increasingly important complement to traditional defenses against ransomware, supply-chain and AI-enabled threats in 2026.
AI's real value to defenders is deciding what deserves human attention first, not doing more work.
MOXFIVE's Ben Harel argues that most organizations do not know with confidence whether their defenses would hold against how attackers operate now.
AI compresses the timeline for attacks teams already understand, letting more attackers move through the kill chain faster, while criminally marketed Dark LLMs extend the ransomware-as-a-service pattern into reconnaissance, phishing, and exploit assistance. Adding more products rarely helps, since the real challenge is clarity about which exposures and control gaps address risk.
AI's greatest value is correlating fragments such as unusual logins, persistence, and lateral movement into prioritized investigations. Digital forensics remains critical because it reconstructs what actually happened, including what controls missed.
Trellix found AI attack services on dark web forums that hand novices APT-grade attack plans.
Trellix researchers found criminal actors selling AI-powered hacking tools on dark web forums, including services named APEX AI, Metamorphic Crypter, and MessiahGPT.
These lower the technical barrier substantially by automating attack planning and defense evasion. Researcher Jambul Tologonov put it directly: someone who would not know where to begin a penetration test can now obtain a prioritized attack plan mirroring APT-grade tradecraft.
Separate Proofpoint research identified indirect prompt injection tools marketed at about $150 a month that manipulate AI agents through instructions hidden in emails, PDFs, and web pages.
Intruder's 2026 index finds weak identity and access management in 80-98% of cloud accounts.
Intruder's 2026 Cloud Security Index reports weak identity and access management affecting between 80% and 98% of accounts across AWS, Azure, and Google Cloud.
Each platform fails differently: AWS shows the highest prevalence of misconfiguration in five of six categories, while Google Cloud has the lowest rates in four.
Weak IAM worsens with scale, hitting 87% of small and midsize businesses, 95% of midmarket organizations, and 98% of large enterprises. Remediation timelines also vary by size, from 8 to 16 days at smaller organizations to a 35-day peak among midmarket firms.
A first-time attendee's account of Black Hat 2026 vendor events, the NOC tour, and community moments.
Cyber Defense Magazine reporter Arya Baviskar recounts a first Black Hat with more than 20,000 attendees and 400 booths.
Highlights include the House of SOC evening hosted by Fig Security and Cribl, a Women in Cyber workout hosted by Seemplicity and Cyera, and a Zluri outing to the Las Vegas Grand Prix track.
A tour of Black Hat's Network Operations Center led by Bart Stump and Neil Wyler revealed that systems flagged more than 9 million potential threats over the week, of which only 308 required blocking.
A talk hosted by Illumio featuring former cybercriminal Brett Johnson underscored that trust remains the first thing an attacker exploits. Across interviews, one theme recurred: AI brings positive possibility, not only risk.
Cybercrime syndicates now out-coordinate law enforcement, which still investigates and indicts in silos.
At Black Hat USA 2026, Carole House of Penumbra Strategies argued that no single actor controls enough of the cybercrime ecosystem to be deterred by law enforcement action.
Threat actors run franchises with divisions of labor, HR functions, and Telegram customer support, while agencies still follow an investigate-attribute-indict cycle against continuously regenerated infrastructure.
House called for a coordinated national strategy modeled on military frameworks, better information sharing across agency priority lists, and international partnerships that impose operational friction.
Huntress researchers add that takedowns matter less now that actors can rebuild infrastructure in moments using AI and vibe coding, making industry-wide collaboration the most promising path forward.
A framework for implementing operational security controls across AI systems using six components.
Organizations should approach AI security as a control-system problem rather than a model evaluation exercise.
The guide presents six operational components: AI system inventory and classification, data use and scope control, identity and access governance, application security for AI systems, monitoring and behavioral oversight, and governance with accountability.
Implementation prioritizes operational controls that prevent AI systems from exceeding defined boundaries. It also extends existing identity governance and behavioral monitoring to address AI-specific attack surfaces such as prompt injection and context manipulation.