Skip to content

Response, New York. Noah Park: Check Point Patches Actively Exploited Management Server Path Traversal. Detection, Dallas. Ava Okello: Report Finds Nearly Half of Deployed Detections Need Attention Before SOCs Can Trust Them.

SOCtember

Always First. Fast SOC News.

Opinion

Commentary from named writers on detection, response, and the tools of daily operations.

Figure 2 from arXiv paper 2610.39351, four pairs of grayscale byteplot images, one pair per benchmark: MalNet-Tiny, MOTIF, Virus-MNIST and Dumpware10.
Figure: Stefan Larson and Kevin Leach, Vanderbilt University, arXiv:2610.39351 (CC BY 4.0), Figure 2, resized and placed on a white background. Each column is one pair of entries from a malware benchmark. Orange frames mark duplicates the audit found; the blue pair is a control that is genuinely distinct.

OPINION · LONDON

Malware Image Classifiers Report 97% to 99% Accuracy, but an Audit of 10 Benchmarks Found as Much as 50.6% of Test Images Had a Copy or Near-Copy in Training, and SOC Teams Should Ask What a Model Was Graded On

Vanderbilt University researchers audited 10 public benchmarks for malware classifiers that read programs as images and found that the standard cleanup, removing files with the same hash, left identical images on both sides of the train and test split. Elena Vos argues that SOC teams should treat any malware model's accuracy figure as a claim about its test set until they know how that set was built.

Elena Vos, Opinion, London

Read the story
Figure 2 from arXiv paper 2610.36693, two line charts of verified joint failures found during an audit.
Figure: Indraveni Chebolu, Seema Panghal, Arnab Mallick and Ch A S Murty, arXiv:2610.36693 (CC BY 4.0), Figure 2, resized and placed on a white background.

OPINION · LONDON

A Detector That Flagged 82.8% of Random Data Corruption Was Blind on 64.7% of the Source and Task Pairs Researchers Audited, and SOC Teams Should Stop Grading Anomaly Detection on Clean Metrics

Researchers at India's Centre for Development of Advanced Computing show that a cross-source anomaly detector with a 3.18% false-positive rate, which flagged 82.8% of random corruptions, was still first-order blind on 64.7% of the source and task pairs they audited. Elena Vos argues that SOC teams should test detectors against an attacker who controls some of their inputs, and add independent checks where cross-checking is weakest.

Elena Vos, Opinion, London

Read the story
Page from Arctic Wolf's 2026 Agentic Security Operations Report showing three headline figures, monthly investigation volume up 21%, alert volume down about 19% and average time to ticket down 36%, beside a pie chart of Arctic Wolf Agentic SOC investigation resolutions: 60% agentic resolution, 38% human-AI resolution and 2% human resolution.
Figure: Arctic Wolf, 2026 Agentic Security Operations Report, page 6, cropped and placed on a dark background. The resolution split is the company's own measure of its own service.

OPINION · LONDON

Arctic Wolf Says Its AI Agents Now Resolve 60% of SOC Investigations Alone, and Buyers Should Ask Every Agentic SOC Vendor How Often Analysts Overrule the Machine

Arctic Wolf's 2026 Agentic Security Operations Report says agents in its managed SOC resolved 60% of investigations with no human involvement and that human-only cases fell from about 81% to 2%. Its most useful page describes an AI that was confidently wrong. Elena Vos argues that buyers should judge agentic SOCs by how often analysts overrule them, not by how much they do alone.

Elena Vos, Opinion, London

Read the story
Figure from the paper showing the experimental setup: an attacker writes a note onto an E01 disk image exhibit, which is ingested into Autopsy 4.23.1 on a Windows analyst workstation; an Autopsy MCP server with 18 read-only tools feeds LM Studio 0.4.23 running the gemma-4-12b model, which the examiner prompts, with optional mcp/filesystem and mcp/fetch servers attached by the examiner.
Figure: Frank Breitinger, Jan Gruber, Christopher Hargreaves, Jan-Niclas Hilgert, Gaëtan Michelet and Mark Scanlon, arXiv:2610.29747 (CC BY-SA 4.0), Figure 4, resized and placed on a white background. The attacker controls only the seized disk image; the tools around it belong to the examiner.

OPINION · LONDON

Planted Text on a Seized Disk Turned an AI Forensics Agent Against Its Examiner, and Incident Responders Should Treat Every Model Reading Evidence as Exposed

Researchers from five European institutions planted plain-text notes on test disk images and got a local AI agent working through the Autopsy forensic tool to relabel photographs, write a false corruption warning to the examiner's machine and contact an outside web address. Responders who point AI at compromised hosts should treat its output as a lead to verify, not a finding.

Elena Vos, Opinion, London

Read the story
Figure from the SecLog-Bench repository showing an example benchmark input and output: a prompt asking the model to predict an attack stage using only visible evidence, an excerpt of audit, application and network log entries, and a model answer of STAGE_UNCLOSED with a four-step reasoning chain citing evidence references E16 to E20.
Figure: Yifan Wang, Dongliang Xu, Xiaoyong Li, Jing Li, Peter S. Pan and Yue Yao, SecLog-Bench repository (github.com/SecLog-Open/SecLog, MIT License). An example prompt, visible log window and model output; the paper notes the example is shortened for readability.

OPINION · LONDON

An AI Analyst That Names the Right Attack Stage for the Wrong Reasons Is Not Ready to Drive Incident Response

A new benchmark from researchers at Shandong University and the University of Technology Sydney found that the best of eight large language models placed only 44.44 percent of attack log windows in the right stage, and that many of its correct answers rested on reasoning the logs did not support. SOCs using AI to scope incidents should grade the evidence chain, not just the verdict.

Elena Vos, Opinion, London

Read the story
Figure 5 from the Alert-Bench paper, an overview of the AIDA triage design: a SIEM alert enters an investigation ledger, then passes through an investigation phase with a Strategist and Executor, a challenge phase with an Adversary, and a synthesis phase in which a Judge applies asymmetric evidentiary thresholds.
Figure: Saimon Amanuel Tsegai, Alex Kantchelian, Danfeng (Daphne) Yao and Peng Gao, Figure 5 of "From Investigation Failures to Reliable SOC Agents," arXiv:2610.10608, Oct. 7, 2026 (CC BY 4.0).

OPINION · LONDON

An Empty Search Is Not Evidence, and AI Triage Agents That Treat It as Proof Will Close Real Attacks

A new benchmark from researchers at Virginia Tech, Google and North Carolina State University found that five common designs for LLM triage agents each missed at least 40.4 percent of attack alerts, often dismissing them after their own searches came back empty. Any SOC handing triage to an agent should require more evidence to close an alert than to escalate one.

Elena Vos, Opinion, London

Read the story
Cover of the Rapid7 Intelligence Quarterly Threat Landscape Report Q3 2026, white headline type over a green dot-matrix field.
Photo: Rapid7 Intelligence, cover of the Quarterly Threat Landscape Report Q3 2026, published Oct. 21, 2026 (fair use).

OPINION · LONDON

When Attackers Exploit High-Severity Flaws in Three Days, the Monthly Patch Cycle Is a Risk Decision, Not a Schedule

Rapid7's Q3 2026 Quarterly Threat Landscape Report finds new high-severity vulnerabilities were exploited an average of three days after disclosure, with 71 percent of exploitation inside the first week and 79 percent of exploited flaws needing no credentials or user interaction. Security operations teams should own the days of exposure, not just the patch ticket.

Elena Vos, Opinion, London

Read the story
SecOps Unpacked hand-drawn ASEF AI SOC Evaluation Framework diagram showing the Screen, Score, Platform and Trust, ROI and Decide stages.
Diagram: SecOps Unpacked / Filip Stojkovski, ASEF framework from the Introducing ASEF post, July 8, 2026 (fair use).

OPINION · LONDON

AI SOC Buyers Need a Lifecycle Score, Not Another Middle-Heavy Demo Checklist

SecOps Unpacked's ASEF framework (July 8, 2026) argues the AI SOC market optimized for triage and investigation demos while detection engineering and response stay thin, and buyers need a vendor-neutral lifecycle profile that refuses blended scores, treats unknown as unknown, and scores trust and feedback loops.

Elena Vos, Opinion, London

Read the story
Cover of the 2026 SANS SOC Survey Insights white paper, A Decade of Evolution in Cyber Defense, by Christopher Crowley.
Photo: SANS Institute, cover of the 2026 SANS SOC Survey Insights white paper, A Decade of Evolution in Cyber Defense (fair use).

OPINION · LONDON

SOC Retention Is an Operations Design Problem, Not a Pay Problem

The 2026 SANS SOC Survey finds meaningful work is still the top retention driver, a 27-point gap between leaders and practitioners on hiring attention, and AI tools used by 79 percent of SOCs but built into defined workflows by only 36 percent.

Elena Vos, Opinion, London

Read the story
Monad Blog cover reading State of Detection Engineering 2026: What the Data Reveals.
Photo: Monad Blog, featured cover for its State of Detection Engineering 2026 coverage of the SANS and Anvilogic survey (fair use).

OPINION · LONDON

Detection Engineering's False Positive Gap Is Ownership, Not Tooling

The 2026 SANS and Anvilogic State of Detection Engineering survey finds most teams track false positive rates while few prioritize cutting them, and vendor default rules remain the main noise source.

Elena Vos, Opinion, London

Read the story
Splunk Enterprise Security product screen with an investigation summary and an isolate-host approve or decline control.
Photo: Splunk. Enterprise Security 8.7 product UI showing the AI SOC Analyst moving from investigation context to a governed response action with Approve or Decline.

OPINION · LONDON

SOC Autonomy Without Governance Recreates the Same Failure Modes AI Was Meant to Fix

Splunk Enterprise Security 8.7 frames the Agentic SOC as a move from AI-assisted investigation to governed autonomous response, where permitted actions, approval gates, evidence, policy, and auditability matter more than buying a stronger model.

Elena Vos, Opinion, London

Read the story

OPINION · LONDON

Alert Fatigue Is a Detection Pipeline Failure, Not an Analyst Character Flaw

LONDON - SOCs still treat missed alerts as a people problem. Detection engineering evidence says most alert fatigue is manufactured upstream in the detection pipeline. Future of SecOps (Aug 2026, Marta K.) described a review that blamed an analyst after a shift queue of more than 800 alerts, including EDR and identity duplicates of one medium-severity credential anomaly. Alert fatigue is volume, noise, duplication, and weak prioritization exceeding review capacity. Intezer (THN Sep 12, 2026): ~16.9M SOC alerts Feb-Jun 2026; ~73k AI-related (0.43%), up 685%; of AI-related, 94.1% benign tool use, 5.8% unsafe/policy, 0.02% confirmed attacks; no confirmed org-agent takeovers. ISACA 2025 via FoS: 55% understaffed; 38% need 3-6 months to fill entry-level. Fix: measure FP/duplicate/backlog/TTA by rule; tune defaults; correlate; severity contracts; rule owners; triage feedback. AI triage helps enrichment but cannot retire orphan rules.

Elena Vos, Opinion, London

Read the story