Regulators in the United States, Canada, the United Kingdom and the European Union now expect organizations to detect intrusions by spotting departures from normal behavior, not only by matching known signatures. Two of these are binding laws that promise fines for organizations that do not comply.
DeepTempo's LogLM is the first in-market solution that updates behavioral anomaly detection through a foundation model that delivers unprecedented accuracy without the extensive initial and ongoing retraining of machine-learning-based systems. You can learn more about this approach to behavioral anomaly detection, developed with the assistance of BNY and other design partners, in the DeepTempo learning center.
This post explains recent changes in regulations in non-legal terms and links to the underlying laws and regulations.
| Jurisdiction | Instrument | Who it covers | What it asks for | Status |
|---|---|---|---|---|
| United States | OMB Memorandum M-21-31, tier EL3 | Federal civilian agencies | User behavioral analytics for early detection of malicious behavior | Binding on agencies since August 2021 |
| United States and Canada | NERC CIP-015-1 (and CIP-015-2) | Bulk electric system operators | Detect and evaluate anomalous network activity from network data feeds | Enforceable October 1, 2028; CIP-015-2 phases from 2029 |
| Canada | OSFI Guideline B-13, section 3.3.2 | Federally regulated banks and insurers | Advanced behaviour-based detection of user and entity anomalies | In force since January 2024 |
| United Kingdom | NCSC Cyber Assessment Framework, principle C2, outcome C2.b | Operators of essential services under NIS; government bodies under GovAssure | Proactive monitoring grounded in normal system behaviour and attacker methods | In use for NIS and GovAssure assessments |
| United Kingdom | Telecommunications (Security) Act 2021 and Security Measures Regulations 2022, regulation 6 | Public telecoms providers | Monitor and analyse Security Critical Functions; investigate anomalous activity | Binding; Ofcom enforces |
| Germany | Telekommunikationsgesetz section 165(3) | Public network operators and service providers with elevated risk | Attack detection systems that continuously and automatically collect and evaluate | Binding; last amended with the NIS2 transposition |
| European Union | DORA, Regulation (EU) 2022/2554, Article 10 | Financial entities and their critical ICT providers | Mechanisms to promptly detect anomalous activities; monitor user activity and ICT anomalies | Binding since January 17, 2025 |
| Eurosystem | ECB Cyber Resilience Oversight Expectations (CROE) | Financial market infrastructures | Baseline normal activity and detect deviations; escalate to correlated, real-time detection | Supervisory expectation since December 2018 |
Context
We have all read about the rise of AI-powered attacks. A few highlights (or lowlights)?
- On November 3, 2026, the UK National Cyber Security Centre published the AI section of its Annual Review 2026. The NCSC emphasizes that AI acts as a force multiplier: it speeds up existing tactics, widens the gap between defenders who use AI and those who cannot, and shortens the window between vulnerability discovery and exploitation. The NCSC also assesses that AI-enabled malware is highly likely to appear in a ransomware attack on UK critical national infrastructure. This is one of many recent assessments leading to a rise in preparation by militaries and regular citizens around the world.
- Attackers already run AI agents. On October 30, 2026, Google Threat Intelligence Group and Mandiant described ARTEX, an open-source multi-agent offensive framework that turns a stated goal into commands across six sub-agents, and its relatives PentAGI, Hexstrike and CAIRN. Since late August 2026, extortion crews and suspected espionage operators have used ARTEX in intrusions against government, technology and manufacturing targets; one actor moved from a first compromised account to deep system access in about four hours, and once in 15 minutes. Google's defensive advice leads with baselining normal authentication, API and traffic patterns and alerting on behavioral and volumetric change.
Here are more details on recent regulatory changes that suggest, or require, the use of a flavor of behavioral anomaly detection in response to AI-powered attacks.
United States
OMB Memorandum M-21-31
Who: Federal civilian executive branch agencies, under Executive Order 14028.
What it says: The memo defines four event logging maturity tiers, EL0 through EL3. The advanced tier, EL3, includes an explicit step to implement user behavioral analytics so agencies can detect malicious behavior early. At minimum the capability must detect and alert on compromised user credentials, privileged user compromise, improper asset access, compromised systems, hosts or devices, and lateral movement by a threat actor.
Why it matters: The memo names the behaviors, not the signatures. Credential misuse and lateral movement rarely trip a known-bad rule; they show up as changes in how a user, host or service behaves.
Source: OMB M-21-31 (PDF).
NERC CIP-015-1 and CIP-015-2
Who: Utilities operating high and medium impact bulk electric system cyber systems in the United States and Canada.
What it says: CIP-015-1 is the first enforceable internal network security monitoring standard. Requirement R1.2 directs utilities to detect anomalous network activity using the network data feeds they collect; R1.3 requires them to evaluate what they detect. The stated purpose is to improve the probability of detecting anomalous or unauthorized network activity inside the perimeter. FERC approved the standard in June 2025 and set it to become enforceable on October 1, 2028. FERC also directed NERC to widen the scope; the resulting CIP-015-2, approved by FERC on August 10, 2026, extends monitoring to electronic access control and physical access control systems in phases that begin October 1, 2029.
Why it matters: The standard uses the word anomalous on purpose. Signature matching on east-west traffic inside a control center misses novel tooling; a baseline of normal flows does not. Utilities are selecting tools now, and flow-log analysis fits the requirement directly.
Source: CIP-015-1 standard (PDF); FERC Order No. 907, RM24-7-000 (PDF); NERC Project 2025-02.
Canada
OSFI Guideline B-13, section 3.3.2
Who: Federally regulated financial institutions, including banks and insurers supervised by the Office of the Superintendent of Financial Institutions.
What it says: Section 3.3.2 on threat detection states that, where feasible, institutions should use advanced behaviour-based detection and prevention methods to detect user and entity behaviour anomalies, alongside emerging internal and external threats.
Why it matters: OSFI names the technique. "Behaviour-based" and "user and entity behaviour anomalies" describe UEBA and its successors, not a signature feed. B-13 has applied since January 1, 2024, so examiners can ask for it today.
Source: OSFI, Technology and cyber risk management (Guideline B-13).
United Kingdom
NCSC Cyber Assessment Framework, principle C2
Who: Operators of essential services regulated under the NIS Regulations, and central government bodies assessed under GovAssure.
What it says: Principle C2, Proactive Security Event Discovery, requires detecting malicious activity even when it evades standard signature-based tools. Outcome C2.b expects organizations to apply their understanding of sophisticated attack methods and of normal system behaviour to monitor proactively for abnormal activity.
Why it matters: The CAF treats signature evasion as the expected case. An organization that cannot show how it would notice an attack with no signature does not meet the outcome.
Source: NCSC CAF, consolidated guidance; GovAssure introduction to the CAF.
Telecommunications (Security) Act 2021 and the Electronic Communications (Security Measures) Regulations 2022
Who: Providers of public electronic communications networks and services, enforced by Ofcom.
What it says: Regulation 6 requires providers to monitor and analyse both access to their Security Critical Functions and the operation of those functions, and to investigate any anomalous activity. The accompanying Code of Practice sets dated tiers for compliance.
Why it matters: The regulation ties monitoring to the functions that keep a network running and makes investigation of anomalies a duty, not a practice.
Source: The Electronic Communications (Security Measures) Regulations 2022.
Germany
Telekommunikationsgesetz section 165(3)
Who: Operators of public telecommunications networks and providers of publicly available services; mandatory for those with elevated risk potential.
What it says: Section 165(3) requires providers with elevated risk potential to deploy attack detection systems (Systeme zur Angriffserkennung). The systems must identify threats through continuous, automatic collection and evaluation of suitable parameters and characteristics, and should be able to stop detected threats and remedy disruptions. The Bundesnetzagentur may set further detail in its catalogue of security requirements under section 167. The provision was last amended on May 12, 2026 as part of Germany's NIS2 implementation.
Why it matters: "Continuous, automatic collection and evaluation" is a statutory description of anomaly detection. A rule library updated weekly does not meet it; a system that evaluates every record against a model of normal does.
Source: TKG section 165 (German).
European Union
DORA, Article 10
Who: Banks, insurers, investment firms, payment institutions, crypto-asset service providers and other financial entities across the EU, plus the critical ICT third-party providers that serve them.
What it says: Article 10 requires financial entities to have mechanisms in place to promptly detect anomalous activities, including ICT network performance issues and ICT-related incidents, and to identify potential single points of failure. Entities must devote sufficient resources and capabilities to monitor user activity, ICT anomalies and ICT-related incidents, including cyber-attacks.
Why it matters: DORA has applied since January 17, 2025. It is a regulation, not a directive, so it binds directly without national transposition, and supervisors can levy fines tied to turnover.
Source: Regulation (EU) 2022/2554 on EUR-Lex.
ECB Cyber Resilience Oversight Expectations
Who: Financial market infrastructures overseen by the Eurosystem, including payment systems, central counterparties and central securities depositories.
What it says: The CROE grades detection across three maturity levels.
- Evolving (baseline): establish a baseline profile of system activities so the institution can detect deviations; monitor user activity, exceptions, faults and security events.
- Advancing (intermediate): continuously monitor the IT environment and detect anomalous activities and events; correlate alerts and anomalies across business units; compare traffic continuously with the expected baseline profile and data flows.
- Innovating (advanced): run correlated analysis of logs, alerts, traffic flows and geopolitical events, and detect and block attacks automatically in real time.
Why it matters: The CROE writes down the ladder most regulators imply: baseline first, then continuous comparison, then real-time correlated response.
Source: ECB CROE (PDF).
Common thread
The eight texts agree on at least four points.
- Signatures are necessary and insufficient. The CAF says so outright; the others assume it by naming anomalies rather than indicators.
- Normal must be defined. CROE, CAF and TKG each require a model of expected behaviour against which to compare. The LogLM delivers this innately.
- The unit of analysis is behaviour over time, for users, entities, hosts and network flows, not single events.
- Detection must be continuous and evaluated. NERC's R1.3 and DORA's resource requirement both make follow-through part of the obligation.
Why anomaly detection
The regulators are responding to the same fact defenders see every day: signature-based detection cannot keep pace with the velocity, volume and novelty of AI-assisted attacks. A signature describes an attack someone has already seen and written up. An adversary who uses AI to generate fresh tooling, rotate infrastructure and move laterally in minutes produces no such history. The only durable reference point is the defender's own environment: what its users, hosts and flows normally do, and when they stop doing it.
We have written about the mechanics of behavioral anomaly detection before; see Behavioral Anomaly Detection: Why Anomaly Detection Earned a Bad Name in Cyber Security and Why deep learning beats traditional ML anomaly detection for today's cyber defense.
Implication for defenders
Classic machine-learning anomaly detection meets these requirements on paper and disappoints in practice for two reasons. The first is burn-in: a model trained from scratch on one environment needs weeks or months of clean history before it can tell normal from novel, and it degrades as the environment changes. The second is noise: a model that only measures distance from normal flags every unusual but benign event, and the SOC learns to ignore it.
A log language model (LogLM) takes a different path on both counts. DeepTempo's LogLM is a foundation model pretrained on security telemetry, so it arrives with a learned sense of how flows, sessions and identities behave and zero-shots in many cases; it can also perform any adaptation to a new environment in days rather than quarters. The LogLM is also a compound model: it performs anomaly detection, then applies classifiers to judge which anomalies are also concerning, mapping them to known tactics and techniques. That combination is why it detects novel attacks zero-shot and why its false-positive rate stays low. Paired with Vigil, the open-source AI SOC, it produces the artifacts these regimes ask for: a baseline, anomaly detections with explanations, and evaluated findings an auditor can trace. Deployments run in the customer's own environment, on-premises, in their cloud or air-gapped, which matters for DORA, TKG and NERC-regulated estates.
Teams mapping their monitoring program to any of the eight instruments above can reach us at deeptempo.ai.
