The Security Boulevard stories enterprise IT leaders actually read.
Volume 343, Issue 2 · 15 articles
The French supermarket fridge attack shows AI-era breaches still enter through old doors: default creds and unpatched devices.
Overnight on October 11 and 12, attackers took control of internet-exposed refrigeration controllers from Danfoss, Copeland and Dixell at at least 350 French supermarkets, pushing fridges to 50C and causing fires and tens of millions of euros in losses.
Claroty had published the flaws two months earlier, including a 10.0 authentication bypass.
AI wasn't confirmed in the attack, but the author uses it to frame 2026, citing IBM's finding that one in four malicious breaches were AI-enabled at an average cost of $6 million and CrowdStrike's report of an 89% rise in AI-enabled adversaries.
Yet the entry points remain familiar, so the fundamentals of exposure, patching and credentials still matter most.
Ungoverned AI agents are shadow IT on steroids; eight controls turn agentic experiments into systems enterprises can trust.
With 90% of enterprises adopting AI agents, the author argues that real AI governance is about control: who can build agents, what data they access, what reaches production and how it is tracked.
Ungoverned deployments bring shadow AI, no audit trail, prototypes accidentally going live and sensitive data exposed to the wrong teams.
Agents also add new attack surfaces in prompt manipulation, retrieval of unauthorized or poisoned content, and actions in real systems.
The piece offers an eight-part checklist covering access permissions, identity, approvals, isolation and auditability, saying that trust, not technology, is the main barrier to scaling.
Dan Lohrmann surveys approaches to agentic IAM, noting there's still no agreed way to manage nonhuman identities.
In his Lohrmann on Cybersecurity column, Dan Lohrmann notes that AI agents may soon outnumber humans in some roles, yet most IAM platforms weren't built to give software a verifiable identity, unleakable credentials and narrow permissions.
Cisco's Matt Caulfield points out that there is no agent equivalent of a human onboarding and background check, even though agents are 'hired' in minutes.
SailPoint stresses that the hardest part is organizational: establishing ownership, required permissions and expiry before agent populations grow too large.
Lohrmann recommends healthy paranoia, full discovery and identification of agents, and operating models that connect AI development, identity and security.
Fenix24 data shows ransomware recovery stalls on Active Directory: 99.2% of victims had no identity recovery plan.
Fenix24's first State of Recoverability report, drawn from more than 500 real ransomware recoveries, found that only four of more than 800 engagements came close to 24-to-48-hour recovery targets, and none reached full operations for weeks.
The bottleneck was identity rather than backups, as 99.2% had no documented plan to recover their identity system and about 20% of the first 48 hours went to establishing one trusted authentication source.
Because attackers who steal the KRBTGT hash can forge Golden Tickets that survive restores and password resets, recovery requires resetting KRBTGT twice.
Boards should demand tested Active Directory recovery plans, not just backups.
AI coding tools produce code faster than senior engineers can validate it, making review the new delivery bottleneck.
ISHIR argues that tools like GitHub Copilot, Claude Code, Cursor and Amazon Q Developer have changed software economics: code can be drafted in minutes, but reviewers still have to understand requirements, check security, evaluate dependencies and confirm architectural fit.
AI lowers the effort to produce code but not the cognitive effort to validate it, so senior engineers become the bottleneck and technical debt can grow faster.
The article urges leaders to measure review capacity rather than output volume, and to adapt review processes, automated checks and team structures so faster pull requests actually lead to better outcomes.
Zero Trust creator John Kindervag argues AI changes attack speed, not security fundamentals: every packet still rides a network.
The author talks with John Kindervag, who coined Zero Trust and now works at Illumio, and who isn't losing sleep over AI because AI can't get above or below TCP/IP: attacks still need network connections and permissions that defenders can monitor and control.
He says much of what AI finds was always there, just undiscovered, and notes that security has stayed at 10% to 13% of IT budgets since 2008 despite the hype.
'All bad things happen inside of an allow rule,' he says, so default deny and least privilege still apply, with AI agents starting at minimal permissions and earning more.
Human expertise, he adds, remains essential.
Overly permissive RBAC, exposed API servers and weak network policies are recurring Kubernetes gaps found in cloud pentests.
Kratikal explains that Kubernetes attack surfaces span clusters, workloads, APIs, identities and cloud integrations, and that many serious gaps need no sophisticated exploit.
Common pentest findings include service accounts with overly broad RBAC rights to read secrets, exec into containers or reach other namespaces, API servers exposed to the internet with weak authentication, exposed dashboards, insecure container settings and missing network policies.
Pentests also show how weaknesses chain together, such as a compromised pod reaching cloud resources.
Recommendations include least privilege for roles and service accounts, regular permission reviews and restricting API server access.
Identity and PAM vendor acquisitions can change pricing, support and complexity, giving customers reason to reassess their tools.
Deals such as Palo Alto Networks buying CyberArk and SailPoint buying Entro Security, along with acquisitions by CrowdStrike, Delinea and BeyondTrust, are consolidating identity security and PAM.
The author says larger vendors offer resources and simpler procurement, but bundled products don't automatically become one platform and can leave separate consoles, policies and workflows.
Licensing may shift to bundles or platform commitments, and acquired products can lose investment and support attention.
With machine identities and AI agents creating new privileged access needs, consolidation is a natural point for customers to re-evaluate whether their PAM solution still fits.
Hidden certificates, orphaned keys and forgotten trust relationships turn routine changes into outages and security incidents.
Few enterprises can say where all the cryptographic identities that let their systems trust one another live.
The gap shows up as an expired certificate on an unknown internal API, a private key outside the approved control plane, a still-trusted old intermediate CA or a machine identity whose owner has left.
Cryptography now spans clouds, Kubernetes, CDNs, gateways, service meshes, SaaS and devices, forming a dependency graph rather than a certificate list.
Without visibility, teams can't predict the blast radius of a change, find owners or prove a key is revoked, so the article urges continuous mapping of cryptographic dependencies.
Bank of America hired 1,000 contractors to fix flaws AI surfaced, raising the question of how remediation must change.
Bank of America brought in 1,000 contractors, on top of about 3,000 staff, to address vulnerabilities in its own and vendor code exposed by advanced models such as Anthropic's Mythos.
As discovery accelerates, the author asks how many people would ever be enough.
Remediation isn't the same as patching, because a patch won't fix excessive privileges, unneeded connections or architectural flaws, so the fix may be revoking access, removing software or isolating assets.
The piece highlights Qualys's AI-driven remediation that goes beyond patching and Aviatrix's segmentation to contain damage while flaws that can't be fixed immediately are addressed.
With security budgets growing only 4%, CISOs must fund AI agent security by reallocating spend rather than requesting new money.
Gartner found that 17% of CIOs had deployed AI agents and 42% more planned to within a year, so fiscal 2027 will be the first full budget year with agents in production for many organizations.
Meanwhile, IANS and Artico Search found that security budget growth slowed to 4%, the lowest in five years, while vendors expect around 12.5% market growth.
LoginRadius founder Deepak Gupta argues that agent security must come from existing spend, so the budget memo should show where money moves from.
He notes that 88% of organizations define privileged users as human only, although 42% of machine identities hold sensitive access.
A primer on SOC 2 Type 2: how it tests controls over time, how it differs from Type 1 and why ongoing audit readiness matters.
SOC 2 Type 2 is an independent attestation based on the AICPA's Trust Services Criteria of security, availability, processing integrity, confidentiality and privacy.
It checks whether a service organization's controls are well designed and operated effectively over a defined period.
A Type 1 report assesses control design at a single point in time, while Type 2 requires evidence that controls such as access reviews were actually performed throughout the period.
The guide explains why SaaS and cloud providers use the report to win enterprise trust, notes that Type 1 isn't a prerequisite, and stresses continuous audit readiness.
How to evaluate AI usage control platforms by the vantage points and data sources they cover, not their feature lists.
Kovrr notes that 'AI monitoring' covers two markets: model observability for in-house apps, and what Gartner calls AI Usage Control, which tracks employees' use of third-party AI and enforces policy.
Picking from the wrong list is the most common mistake, and because network, browser, endpoint and identity vantage points each see only part of an interaction, architecture matters more than features.
The guide lists eight surfaces a platform should cover, including agents and MCP servers, LLM telemetry and DLP.
It advises buyers to count live integrations and to ask how each source collects data, attributes identity and refreshes.
In an agentic SOC, AI handles routine triage while Tier 2 analysts become threat hunters and governors of AI agents.
The author walks through a Tier 2 analyst's day in an agentic SOC, where AI agents handle triage, correlation, evidence gathering and enrichment while a QA agent checks their work.
The day starts with agent performance metrics rather than an alert backlog.
Suggested targets include an autonomous closure rate of around 85% (97% in some environments), investigations starting within 5 to 7 minutes, multi-vector triage in 8 to 24 minutes and about 15% of cases escalated at a 90%-plus true-positive rate.
Human-in-the-loop approvals, daily QA and continuous agent tuning keep the automation accountable.
A guide to threat intelligence platforms judged by how fast they turn an IOC into a confirmed investigation and response.
The guide argues that an IOC match is a lead, not proof of compromise, so a threat intelligence platform should be judged by how quickly it links indicators to your own telemetry and a response decision.
Citing SANS, where 73% name false positives as their top detection problem, and IBM's $4.44 million average breach cost, it advises hybrid SOCs to weight IOC monitoring and SIEM, XDR and SOAR integration most heavily.
It also recommends distinguishing standalone from embedded intelligence.
Written from Seceon's perspective, it ranks Seceon TI360 first, also covers Recorded Future, Google, Anomali and ThreatConnect, and recommends identical IOC-to-response tests for every candidate.
More from Security Boulevard · All IT section newsletters · This week