🛡️ IT Operations & Security
AI for keeping digital systems running, observable, and secure. One of the most mature domains: log analysis, threat detection, and automated remediation are established or good practice. AIOps and SIEM are mainstream. Bleeding-edge frontiers include autonomous incident response and AI-driven penetration testing. Five practices are actively advancing; the rest are holding steady at good-practice level.
The Headline
AI now finds software flaws and diagnoses outages faster than anyone can verify its work. The bottleneck has moved to checking and fixing, and attackers are using the same tools.
The Picture
AI-powered threat and anomaly detection now ships as standard in every serious security and IT operations platform, so having it no longer sets you apart. What matters is what happens after an alert fires. Most organizations let AI investigate and draft, then keep a person approving every change. In Dynatrace's own survey, 87 percent use AI agents (software that acts on its own) in IT operations, but only 12 percent run fully autonomously. A small group lets machines act unattended on narrow, well-understood fixes. Qualys says 40 million of its 150 million patches in the past year ran with no human involved, with a rollback rate under 0.1 percent. Everyone else is up against attackers who now use the same AI tooling, at a speed that manual processes cannot match.
This Fortnight
A compromised open-source package shipped on November 9 with every standard trust signal intact. The bee-js release went out through the project's own trusted publishing workflow with valid signed build records. A trojanized desktop app from the same project was notarized by Apple and stayed available for about 8.5 hours, reaching users through its auto-updater. Any check that only asks whether a release is signed would have let it through, so ask your engineering teams what else they verify before a dependency reaches production.
Only 8.4 percent of disclosed AI-found vulnerabilities have been patched upstream. One tally of 6,157 disclosed findings counted 516 fixes. Separately, security leaders at large companies say validating a single critical finding takes more than a day. Adding more AI scanning will grow your backlog faster than your fixes, so put the money into validation and patching capacity rather than more discovery.
Vendors' accuracy claims for AI outage diagnosis got bolder, with AWS citing a customer at 90 percent first-pass root-cause accuracy. The figure comes from Deriv, an online broker, and AWS gives no sample or method for it. An independent benchmark finds agents regularly invent causes that aren't there, and Thoughtworks calls AWS's agent early and advises checking every finding it produces. Treat these tools as a way to speed up your engineers, and before paying for autonomy, ask vendors to show results on your own past incidents.
A practitioner's reconstruction described an account-lockout playbook that ran zero times in five months without anyone noticing. Six links in the automation had broken silently, including a revoked access token belonging to someone who had left, so containment fell back to a person working by hand. An automated response that is never tested only works on paper. Require a scheduled test incident that must trigger it.
Attackers are now running the open-source AI hacking tools that defenders use for testing. Google's Mandiant team has found frameworks such as ARTEX in live intrusions since late August, ranging from extortion gangs to suspected state espionage. In one tracked campaign, attackers reached deep system access in about four hours. Your response plans should assume intrusions unfold in hours, not days.
Coming Up
On December 7, new US federal rules replace uniform patch deadlines with risk-based tiers. The Cybersecurity and Infrastructure Security Agency's directive gives agencies three days to act on the highest-risk exposures. A companion FedRAMP notice (the federal cloud security program) applies risk-based remediation to federal cloud providers. Suppliers to government will see these expectations passed down to them, so check now whether your team can rank flaws by real-world exploitability rather than severity score.
Splunk ends support for its standalone user-behavior analytics product on January 31, 2027. Splunk stopped selling the product in December 2025 and now offers the capability only in its premium security tier, part of a wider move that folds identity anomaly detection into larger suites. If you run it, decide this quarter whether to upgrade, migrate or replace it, because new behavioral baselines take weeks to months to learn.
Microsoft is extending its data-loss controls to Box, Google Workspace, Dropbox and Salesforce against a January 2027 deadline. A competitor's reading of Microsoft's documentation says the current preview covers only stored data behind a connector, with no in-the-moment user warnings and no test mode. Before retiring another tool on the strength of that roadmap, test what the preview actually catches in your environment.
What's Hard About This
Fewer organizations are fully patched, even as AI finds more flaws. Verizon's 2026 breach report, as cited by CyberProof, puts full patch saturation at 26 percent, down from 38 percent. Discovery scales with computing power, but a fix still needs an owner, testing and a maintenance window, and buying more tools does not remove that limit.
AI alert triage can look efficient while quietly missing real attacks. When researchers replayed a set of past alerts, each of five AI triage approaches missed at least 40.4 percent of the alerts tied to attacks, often dismissing them when a search came back empty. A shorter queue does not mean a safer system, so judge these tools by attacks contained, not tickets closed.
Valid signatures, logins and sender checks no longer prove that something is safe. KnowBe4 reports that 58.1 percent of phishing emails that get past email gateways still pass the standard sender-authentication check, and session hijacking was the most common way into accounts in Prophet Security's customer environments. Controls that assume a trusted source is safe need a second layer that watches how the account or software actually behaves.
Practices in this Domain (20)
Read the full technical briefing (1,483 words) →
Where AI Stands in IT Operations & Security
Detection is no longer the interesting question in IT operations and security. Behaviour-based malware detection, log anomaly detection and alert correlation ship as standard in every serious platform. The commercial leaders are large and still growing: CrowdStrike reported $5.25B in ending ARR for fiscal 2026, and Tenable holds 24.6% of its market for an eighth consecutive year. Independent testing puts enterprise endpoint protection in the mid-to-high nineties. The domain's maturity now depends on what happens after an alert fires: whether anyone validates it, whether the diagnosis is right, and whether a machine is allowed to act on it. On each of those counts the evidence is thinner and more contested than the product roadmaps suggest.
Momentum is building in agents that investigate, draft and propose, and it stops at the point of acting. In Dynatrace's own survey, 87% of organisations use AI agents in IT operations, but only 12% report fully autonomous operation, and humans still verify 66% of AI decisions. The named successes keep a human at the gate. Daimler Truck's self-healing SRE agents cut resolution time by about 88% across 36+ applications, yet every patch goes out as a pull request for an engineer to approve. Halter's Claude-based agent had 82% of its 431 generated pull requests merged. Unattended remediation works only where failure classes are narrow and well understood. Qualys says 40 million of its 150 million patches in the past year ran fully autonomously, with a rollback rate under 0.1%. Outside that slice, the benchmarks are sobering. Top agents resolved 13.8% of IBM's ITBench SRE scenarios, and the best root-cause accuracy on ORCA-bench is 25.3% on medium tasks and 10% on hard ones.
What sets this domain apart from its neighbours is that its progress is measured against an opponent who uses the same tools. Google's threat intelligence teams and Mandiant now find open-source offensive agent frameworks such as ARTEX, Hexstrike, CAIRN and PentAGI in live intrusions. Anthropic documented a Midnight Blizzard-linked actor whose workflow automatically rebuilt its toolkit whenever it was detected. Malware such as BONZAI_COBUCH injects fabricated messages to mislead AI triage agents. Advances in finding flaws are shared with attackers, while the work of fixing them is not. Anthropic's Project Glasswing found more than 10,000 high or critical vulnerabilities in about a month, and by 22 May 75 had been patched. The defender's bottleneck has moved downstream, to validation, remediation and governance, and capacity there is growing far more slowly than discovery.
What's New, 2026-10-30 to 2026-11-13
No practice changed position this fortnight. What the evidence did was sharpen the gap between what vendors claim and what independent checks find. In incident diagnosis, vendor accuracy figures got bolder. AWS published Deriv reporting 90% first-pass root-cause accuracy and a cut in MTTR of at least 50%, with no sample or method given. NeuBird, whose platform went GA in mid-October, claims root cause in under five minutes with 94% accuracy. Traversal raised $120M at a $1B valuation on monthly investigations rising from 513 to nearly 73,000, even though ORCA-bench, which Traversal researchers helped build, finds hallucinated causes in 7–40% of agent reports. Thoughtworks' Radar called AWS DevOps Agent on live alarms "early" and advised validating every finding. PagerDuty's Paige agent went GA, drafting post-incident reviews, with TELUS named as a user. A 260-run lab test in the same period found agent-written incident reports claiming more than their logs showed, and the pass rates depended on which model did the judging. A CloudRaft buyer's guide found most AI SRE tools still default to read-only. It names Resolve AI, Komodor, Traversal and NeuBird as the handful that act on production.
On the security side, the discovery-to-fix funnel got a hard number. Of 6,157 disclosed AI-found vulnerabilities, 516 were patched upstream, or 8.4%. HackerOne's survey of 111 leaders puts the average time to validate a critical finding at 30 hours. IBM and Red Hat's Lightwell, which pairs AI with human validation, has repaired more than 400 previously unknown Java library flaws. OpenAI self-reported patching 948 Linux privilege-escalation bugs for $1.2M in API costs. Offensive tooling spread further: Unit 42 counted 609 internet-facing ARTEX hosts and 1,323 running CyberStrikeAI, even as Horizon3 reported NodeZero in use at more than 7,000 organisations. On 9 November a compromised bee-js npm release shipped through the project's own trusted-publishing workflow with valid SLSA provenance, and the trojanised desktop app was notarised by Apple. Smaller signals pointed the same way. A Sentinel account-disable playbook that had run zero times in five months, leaving containment to take 53 manual minutes. Only 4 of 25 organisations interviewed had exercised an agent incident scenario. And a Cybersecurity Insiders summary found 17% have fully implemented Universal ZTNA, against 82% who call it essential.
Key Tensions
Discovery at machine speed, remediation at human speed. AI has multiplied the number of flaws found without changing how many get fixed. Only 8.4% of disclosed AI findings in one funnel were patched upstream, validating a critical finding takes 30 hours on average, and nearly 70% of HackerOne's respondents say validated findings enter the backlog faster than they can remediate them. Verizon's 2026 DBIR, as cited by CyberProof, puts full patch saturation at 26%, down from 38%.
Agents may look but not touch. Operators readily let AI investigate and rarely let it act. Grafana's survey found 85% comfortable with read-only AI investigation against 52% for autonomous remediation, PwC found only 22% would let agents execute defensive manoeuvres without approval, and the Cisco-commissioned Omdia figure of 51% running agentic AI in production shrinks to 18% when the question is remediation. The deployments that work keep autonomy to failure classes such as restarts, rollbacks and allow-listed fixes.
Self-reported accuracy against independent benchmarks. Vendors publish accuracy figures of 90% and above for root-cause analysis and DLP precision, mostly without method or sample. Independent tests are harsher. Five LLM triage approaches each missed at least 40.4% of attack alerts in a 1,247-alert replay, DLP blocked only 16% of confirmed losses before the data left in a CSA survey, and an AI phishing detector fell from 99.78% in controlled tests to 77.8% on unfamiliar mail.
Adversaries run the same agents. The open-source pentesting harnesses defenders evaluate are now staple attacker kit. Microsoft tied Artex to Storm-0940 activity reaching more than 2,250 cloud tenants, and Wiz tracked one campaign reaching deep access in about four hours, in one case 15 minutes. Attackers also target the defender's AI directly, through malware that manipulates triage agents and AI-assisted labs testing more than 70 evasion techniques against live EDR stacks.
Signed, authenticated and still hostile. Signals that once meant a source could be trusted no longer do. The bee-js release and LiteLLM's second compromise both carried valid provenance from compromised release paths, KnowBe4 finds 58.1% of the phish that gateways miss still pass DMARC, and session hijacking was the most common way into accounts in Prophet Security's customer environments. Detection that relies on credentials, signatures or sender reputation is being beaten by attacks that arrive with all three.
Top 10 Evidence Items
- Finding Went to Machine Speed. Fixing Didn’t. (opinion) — Gives the hard number behind the discovery-to-fix funnel the briefing leads with: 8.4% of AI findings actually patched. https://phoenix.security/ai-vulnerability-discovery-remediation-bottleneck/
- AI Ambition Is Outpacing Operational Proof: Dynatrace Warns of Growing Observability Gap (adoption-metric) — Source of the headline 87%/12%/66% stats that anchor the claim that agents investigate but rarely act unsupervised. https://enterpriseitworldmea.com/ai-ambition-is-outpacing-operational-proof-dynatrace-warns-of-growing-observability-gap/
- Traversal Raises $120M Series B at $1B Valuation to Expand AI SRE (news-coverage) — Shows the vendor-growth-versus-independent-benchmark tension directly: a $1B raise against ORCA-bench's own hallucination findings. https://www.unite.ai/traversal-raises-120m-series-b-at-1b-valuation-to-expand-ai-sre/
- Deriv achieves 90% first-pass root-cause accuracy with AWS DevOps Agent (case-study) — Exemplifies the bold, method-free accuracy claims the briefing contrasts with independent testing. https://aws.amazon.com/solutions/case-studies/deriv-aws-devops-agent/
- Malicious bee-js npm Release and Swarm Desktop Releases Deliver SleepyDuck Ethereum C2 Backdoor (case-study) — Demonstrates the 'signed and still hostile' tension: valid SLSA provenance and Apple notarisation did not stop a backdoored release. https://socket.dev/blog/bee-js-compromise
- ARTEX and Friends: Enter Agentic Offensive Frameworks | Google Cloud Blog (industry-report) — Direct evidence that attackers run the same agentic offensive frameworks defenders are still evaluating. https://cloud.google.com/blog/topics/threat-intelligence/artex-and-friends-enter-agentic-offensive-frameworks
- A thousand fewer paths to root (case-study) — A rare large-scale AI-driven fix at cost and effort that still concedes validation and prioritisation remain the bottleneck. https://openai.com/index/a-thousand-fewer-paths-to-root/
- AI cyber resilience and the trust gap (industry-report) — Quantifies how unprepared operations teams are for agent-specific incidents, undercutting claims of operational maturity. https://siliconangle.com/2026/11/08/ai-cyber-resilience-and-the-trust-gap/
- AI Security Accountability: Who Owns the Call? | HackerOne (industry-report) — Puts a concrete 30-hour validation cost on each finding, explaining why discovery has outpaced remediation capacity. https://www.hackerone.com/blog/ai-security-accountability
- How Cursor Built AI-Native Security Operations | Artemis (case-study) — A named, high-autonomy deployment that still keeps a human gate on any action, illustrating the 'look but don't touch' pattern. https://artemissecurity.com/customers/cursor