The security stories enterprise IT leaders actually read.
Volume 343, Issue 3 · 20 articles from 3 issues
A buyer's guide to AI security platforms that unify EDR, SIEM and XDR, with compliance evidence as a key criterion.
This comparison argues that the best AI cybersecurity platform puts endpoint, log, network and identity data on one data layer, automates response with guardrails and generates compliance evidence from the same telemetry.
It cites IBM's average breach cost of $4.44 million and a SANS finding that 73% of organizations rank false positives as their top detection problem.
Written from Seceon's perspective, the guide ranks Seceon OTM first for compliance-focused buyers and also evaluates Microsoft, Palo Alto Networks, CrowdStrike Falcon and SentinelOne.
It advises testing every shortlisted platform against the same proof-of-concept scenarios.
Exploitation overtook stolen credentials as the top breach vector, and AI discovery is flooding teams with more patches than they can apply.
Verizon's 2026 DBIR found that vulnerability exploitation became the leading initial access vector for the first time in 19 years, at 31% of breaches versus 13% for credential abuse, even before AI discovery reached scale.
Claude Mythos then found a 27-year-old OpenBSD bug and more than 23,000 potential flaws across 1,000 open source projects.
Linux disclosed 2,308 CVEs in the first half of 2026, FIRST forecasts about 66,000 CVEs this year, and median remediation time rose from 32 to 43 days.
Because most CVEs are never exploited, the author argues for prioritizing by real exploitation and exposure rather than trying to patch everything.
AI's real effect on attacks is tempo in reconnaissance and post-theft analysis, so defenders must speed up triage.
This opinion piece separates what AI can do in a lab from what attackers actually do at scale.
AI-written malware remains immature and detectable, and the bigger change is in reconnaissance and post-theft analysis, where a model can sift stolen data in any language in an afternoon instead of months.
Attackers also self-host open-weight models, with around 175,000 exposed Ollama hosts operating outside provider abuse monitoring.
Rather than simply generating more findings, the author advises automating triage, enrichment and investigation, and treating the autonomous SOC as a maturity model with humans on the loop for key decisions.
CrowdStrike documents threat actors running agentic AI campaigns and finds token budgets, not engineering skill, limit AI-run intrusions.
CrowdStrike Intelligence reports several 2026 incidents in which threat actors used AI agents to make decisions and execute commands, not just to write tooling.
The report profiles four actors, two using sophisticated AI harnesses and two using simple ones; even the simpler setups breached numerous organizations.
CrowdStrike concludes that the main barrier to scaling AI-orchestrated intrusions is token budgets rather than engineering complexity.
It warns that low-skilled attackers will become more dangerous over the next year and that businesses must modernize defenses to keep up with faster, longer-running automated attacks.
Ransomware crews target hypervisor, storage and backup consoles that often lack MFA while perimeter spending soars.
Ransomware crews get paid when victims cannot restore, which means controlling hypervisors, storage arrays and backups.
The author notes that these consoles often require only a password, sometimes a shared one, because MFA programs were scoped by headcount and by what insurers asked about first.
With Palo Alto Networks reporting that 65% of initial access is identity-based, attackers arrive with valid credentials and the perimeter treats them as employees.
The piece argues that security budgets have their priorities backwards, leaving the systems needed for recovery less protected than the perimeter.
Proofpoint says AI-generated attack chains are making some threat actors worse, and top actors largely avoid AI.
Countering headlines claiming that most ransomware now uses AI, Proofpoint researcher Selena Larson told Dark Reading that AI is making a lot of cyber activity dumber.
Proofpoint highlighted TA4903, a once-sophisticated cybercrime actor whose campaigns have degraded as it leaned on AI-generated attack chains.
Larson pushes back on the narrative that AI is making threat actors better and says the most capable actors are largely staying away from it.
The article notes that AI remains useful to attackers for research and other support tasks, but cautions against vendor and media hype.
IBM says open-source defenders should triage by real exposure and make sure fixes reach neglected components.
An IBM report argues that the open-source community can only keep up with vulnerability volume by triaging the most serious flaws and distributing fixes effectively.
Defenders should ask when a vulnerability actually makes a system vulnerable, since severity scores miss context and AI agents can chain low-severity bugs into damaging intrusions.
Teams need to know which systems an attacker could reach and what mitigations would stop them.
Drawing on its experience running a vulnerability clearinghouse, IBM also stresses ensuring that fixes reach under-attended components.
When AI-driven NDR triggers automated response, teams must retain enough raw evidence to justify the decision later.
The explainer opens with an automated response that isolated a production segment on a behavioral model's verdict, after which reviewers could not reconstruct which flows triggered it because raw flow retention had expired.
Signature alerts carry their own justification by naming the rule and matched values, but a behavioral score names neither.
The piece argues this is a retention policy that was never recalibrated, not a logging failure.
It describes which detections and irreversible actions still require deterministic, reproducible evidence and how to align retention with automated response.
Ransomware crews increasingly use physical and psychological threats against staff and families, a Semperis survey finds.
Semperis incident responder Tim Beasley describes receiving an unsigned package of Russian souvenirs while investigating Midnight Blizzard, an example of intimidation spilling into real life.
A global Semperis survey found that 37% of ransomware attacks involve psychological or physical threats.
The article links the trend to The Com, a loose network whose young members engage in hacking, swatting, extortion and violence, and notes that Europol flagged 4,340 URLs for removal in a summer crackdown.
Responders say violence as a service is now being used to pressure victims into paying.
Censys found internet-exposed AI software grew 63% in nine months while exposed ICS and legacy systems barely moved.
Censys counted AI-related software across 44 products on 267,915 internet-exposed IP addresses in July 2026, up 63% from 164,088 nine months earlier.
n8n accounts for 113,926 of those hosts, and 76.1% of versioned n8n servers run a release affected by at least one vulnerability scored 9.0 or higher while storing credentials for connected services.
OpenClaw went from two exposed control panels to 85,478 in three months.
By contrast, exposed industrial control systems grew about 4% in two and a half years, and old versions of Jenkins, Cleo MFT, Telerik and SharePoint barely changed share, suggesting abandoned instances.
Bitsight found a criminal tool that runs autonomous hacking on stolen OpenAI, Anthropic and Google API keys.
Bitsight tracked thousands of posts on 16 underground forums from February to August 2026 and found mentions of AI agents, indirect prompt injection and MCP up 67%.
One seller is marketing MONST3R-AI, an autonomous penetration testing platform where buyers paste stolen AI provider keys, set a budget and point it at targets, so the key owner pays and provider safeguards are sidestepped.
The seller claims 1,072 findings across 199 hosts, which Bitsight could not verify.
Companies that leak AI keys and teams that install agent plugins from public registries are both exposed.
Researchers show a cheaply backdoored open-weight model passes evals, then exfiltrates credentials when triggered.
Schneier highlights research showing that any open model whose weights have been edited, including fine-tunes, merged adapters and abliterated builds with refusals removed, can carry a backdoor.
The researchers built one cheaply from an open base, a public dataset and a rented GPU, using a trigger that fetches a remote payload so behavior can change months after deployment.
When scaled to 7B and run through OpenAI's Codex CLI, the model answered normal requests but exfiltrated project credentials when triggered.
Schneier notes that AI systems are vulnerable to the same supply chain attacks as regular software.
Identity infrastructure is a prerequisite for ransomware recovery, and identity and BCDR teams need to plan it together.
Data and application recovery plans are useless if staff cannot authenticate because the identity infrastructure is down, as the Maersk NotPetya incident showed when a single surviving domain controller made rebuilding possible.
Omdia research shows the average enterprise uses about 1,200 applications and frequently runs multiple identity providers, including CIAM and merger-driven workforce IdPs.
Identity teams also face configuration drift and admin errors that can halt operations.
The article argues that identity resilience requires restoring components in the right order with their upstream and downstream dependencies, not just backing up the IdP to meet RTOs and RPOs.
Specialty carriers now sell AI coverage while major insurers add AI exclusions, so firms should review policies before year-end renewals.
Some insurers, including Armilla, Coalition, Cowbell, Munich Re and Beazley, offer AI-specific endorsements or policies, while carriers such as Chubb, Travelers and AIG are adopting Verisk/ISO exclusions for generative AI in general liability policies.
Many enterprises wrongly assume existing cyber or liability policies cover AI losses.
The article lists steps to qualify for coverage: review current policies, assess guardrails for agent attacks, check supply-chain contracts, give AI apps least privilege, and monitor agents like service accounts.
Underwriters assess controls holistically, and buyers should look for explicit wording on model outputs, non-malicious failures and bias exclusions.
Attackers hijacked outdated refrigeration controllers at French grocers, spoiling food while masking it as defrost cycles.
Hackers broke into commercial refrigeration controllers at major French grocery chains overnight, turning fridges and freezers into heaters reaching 50 degrees Celsius, spoiling stock and reportedly causing small fires.
To avoid detection, the attackers extended defrost cycles so the heating looked like routine maintenance.
Hundreds of stores lost tens to hundreds of millions of euros.
The article highlights that grocers relied on third parties to manage outdated OT firmware on internet-reachable controllers, a lesson in OT asset ownership and patching across retail.
AWS's Rudra Mitra says frontier models like Claude Mythos force security to respond at machine speed.
Computer Weekly's Cliff Saran talks to Rudra Mitra, AWS vice-president of security services, about cybersecurity in the 'post-Mythos era'.
Anthropic's Claude Mythos, unveiled in April 2026, showed how quickly frontier AI can find and weaponize vulnerabilities, collapsing the gap between discovery and exploitation.
Mitra argues that the old approach of collecting telemetry for dashboards can't keep up, so defenses must also move at machine speed.
He describes AWS Continuum, a model-agnostic tool that automates vulnerability scanning and patch prioritization across the code vulnerability lifecycle, alongside layered defenses and responsible disclosure.
Major security incidents including F5 rootkit, Cisco vulnerabilities, and AI agent exploitation reported.
This week's cybersecurity roundup highlighted critical threats across enterprise infrastructure.
Attackers deployed a Linux rootkit on F5 BIG-IP APM devices that hides web shells in memory, while nation-state and ransomware actors actively exploited authentication bypass flaws in Cisco Firewall Management Center.
Additional concerns include AI agents automating cyberattacks with minimal human involvement, breaches affecting millions including student data and crypto wallets, and emerging phishing techniques leveraging browser-based fake login pages and social engineering against personal devices.
The arrest of a well-known ransomware negotiator in the ShinyHunters probe shows the need for incident vendor oversight.
Edward Dubrovsky, a prominent negotiator whose former firm Cypfer sponsored a Philadelphia cyber summit days earlier, faces federal extortion charges tied to the ShinyHunters investigation.
The core allegation is that a man organizations paid to negotiate with extortionists may have been working with them.
The author frames this as a program failure rather than individual misconduct, since ShinyHunters has run credential theft and SaaS extortion for years, reportedly earning over $70 million.
The piece urges organizations to vet and monitor incident response and negotiation vendors as rigorously as any other high-risk third party.
Instead of caging general-purpose AI agents, build small single-purpose agents with least privilege.
Mark Baggett argues that the industry is approaching AI security backwards by wrapping powerful general-purpose agents in containers, guardrails and human approval prompts.
If model makers struggle to control behavior, he questions whether users can reliably contain them, and he notes that models get retired and replaced, invalidating prior testing.
He recommends using capable agents to build smaller, single-purpose agents that have only the functions a task needs, such as reading logs from one directory.
This approach applies least privilege and separation of duties to agent design to limit the blast radius.
Experts urge moving from vendor severity scores to real exploit data and environmental context as CVE volumes explode.
Record Patch Tuesdays and AI-driven discovery are overwhelming vulnerability programs, and one organization's system collapsed after September's release because fewer than 1,000 CVEs spread across tens of millions of assets.
Root Evidence's Jeremiah Grossman says only 1.4% of CVEs are exploited and the number adversaries actually exploit has stayed flat.
Empirical Security's Ed Bellis and Jay Jacobs say teams suffer because they prioritize manually with thin data.
The experts recommend KEV catalogs, exploit evidence and environment-specific context over theoretical scores.